🚀 No GPU? Run it on Runpod Serverless

ForgeHub — desktop app (macOS / Windows / Linux) that drives this whole stack: recipe UI, mmh3/ wildcards, reference image + video upload, style LoRA selector, QwenVL prompt enhancer, and a chat that writes MiniMax H3 spec prompts for you.

⬇️ Download ForgeHub · ⚡ Deploy the worker on Runpod

Deploy the endpoint → attach a ≥160 GB network volume → paste the endpoint ID into ForgeHub settings → generate. All models auto-download on first boot.

Prefer a full ComfyUI pod instead? Same stack, one click:

Want both stacks (video + Qwen Image 2.1 image gen/edit)?

⬆️ 2026/10/23 UPDATE ⬆️

🎬 MiniMax H3 Director — Two New Workflows

The pack adds the Director pair — timeline-driven multi-scene videos with audio, powered by the new MiniMaxH3-Director custom node — huchukato/ComfyUI-MiniMaxH3-Director-Enhanced. Not yet in the Manager registry — install via git and restart ComfyUI:

cd ComfyUI/custom_nodes
git clone https://github.com/huchukato/ComfyUI-MiniMaxH3-Director-Enhanced

Director-Native (MiniMaxH3-Director-Native-Qwen3.5.json): character sheet + ideas one per line (each line = a segment) → the Director compiles the full timeline prompt itself. Auto-Analyze builds the subject description from the sheet; the built-in QwenVL Enhance expands each segment. The compiled prompt is what runs — visible in the ShowText node
Director-Storyboard (MiniMaxH3-Director-Storyboard-Qwen3.5.json): generate up to 3 scene stills with QwenImageEdit21-Wildcards-Qwen3.5.json (preset IMG › Storyboard), load them chronologically — the in-graph QwenVL turns stills + idea into the final R2VA prompt (external_prompt), the Director renders it with the stills as references
Quality preset by default: both Directors run the 20-step quality recipe (res_multistep + simple) — at 6 turbo steps reference adherence drifts (subject out of frame) and audio doesn't converge (garbled speech). The config dropdown offers r2va_singularity_turbo (euler + beta + 6 steps + ref2v turbo LoRA) for fast drafts without identity/speech-critical content
References switch (Native): REF2VA (Refs ON — the sheet drives identity, default) / OFF (text-only, fl2va path)
Style dropdown like Singularity: [MATCH REF1] follows the reference's art style, or force one (cinematic, anime2d, claymation, …) — expands via the mmh3/style/ wildcards
Requires QwenVL-Mod ≥ 2.12.2 (GGUF import fix — nodes register cleanly even without llama-cpp-python)

🪄 Director Canvas — Auto-Enhance + Wildcards

Auto-enhance at queue: queueing a Director workflow expands every segment still holding its simplified idea through the built-in QwenVL — the same thing the serverless handler does per job. Editing a segment's text re-enhances it on the next queue
Enhance button fixed: it now actually appears on the Segment Prompt panel (single segment selected) for one-off manual re-runs
Wildcards on canvas: the enhance route expands __mmh3/...__ tokens in ideas and the global prompt — no WildcardProcessor node needed
Requires MiniMaxH3-Director ≥ 0.3.8

💬 Qwen Chat — Storyboard Flow

With a Qwen Image 2.1 workflow open and a reference attached, ask for a storyboard: the chat invents 3 chronological still prompts (one location, evolving light, one gesture each) + the timestamped Director idea, and offers Generate Still 1/2/3 buttons that set the prompt and queue one still per turn
Chat reply budget doubled: max tokens 1024 → 2048
Requires QwenVL-Mod ≥ 2.12.2

🖼️ ForgeHub v0.1.18

Both Director workflows ship in the ForgeHub form: charsheet upload, ideas field, style/config dropdowns, resolution picker, and the prompt trace showing the compiled prompt after every run. Download ForgeHub

⬆️ 2026/10/07 UPDATE ⬆️

⚡ Singularity — One Fused Checkpoint for R2VA

The R2VA stack now runs on WarmBloodAban's Singularity ref2va fusion (Minimax-h3_Singularity_ref2va_Pruned_v1.3_int8.safetensors, ~21 GB) — WarmBloodAban/Minimax-h3_Singularity:

One checkpoint for T2V / I2V / R2V / V2V — reference conditioning is fused into the weights
⚠️ Do NOT attach the rank-256 reference LoRA — it's baked into the fusion and breaks the style. Native R2VA on the plain ref2va checkpoint still needs it, Singularity does not
Up to 3 reference images + 1 optional reference video; a single image works fine and behaves like I2V
Optional style LoRA slots baked into the workflows: h3-realism-people-t2v-i2v-r2v and h3_character_swap_pro4500_1000 (set strength 0 / node off to disable)

🎚️ Correct Turbo Settings — `euler` + `simple`

Per the official lightx2v spec, the 8-step turbo LoRAs want euler sampler + simple scheduler with MiniMaxH3SigmaShift at 6/3 (768p training grid). er_sde/beta pairings floating around are for other LoRAs — don't use them here.

🚀 ForgeHub + Runpod Serverless

The whole MMH3 stack now also ships as a serverless worker on the Runpod Hub — zero local GPU required:

Hub listing: huchukato/runpod-minimax-h3 — one-click endpoint deploy, attach a ≥160 GB network volume and the models auto-populate
ForgeHub desktop app drives it: recipe UI, wildcard quick-insert (vid/ TagForge wildcards), reference image/video upload, style LoRA selector, QwenVL prompt enhancer toggle, and a recipe-aware chat that writes MiniMax H3 spec prompts for you
Presets shipped in-app: r2va_singularity (20 steps res_multistep/simple, shift 4/3) and r2va_singularity_turbo (8 steps euler/simple, shift 6/3, ref2v turbo LoRA)

🧠 Prompt Enhancer — Two Tips

Hand-written spec prompts (the subject_definitions/detailed_description format): disable the enhancer — passthrough keeps your exact structure; the enhancer preset rewrites it and can drop your <Picture N> bindings
keep_model_loaded keeps the GGUF enhancer resident between runs — worth it on big-VRAM cards if you use it every job

⬆️ 2026/10/02 UPDATE ⬆️

⚡ Workflow Stack Cleanup

Comfy Kitchen Attention is now wired directly into every safetensors workflow between LoraLoader and BlockSparseAttention
SeedVR2 removed completely from the MiniMax H3 workflow pack; TensorRT upscaling now receives frames directly
The redundant I2VA-V2VA workflow has been retired
The pack now ships 4 workflows: T2VA, I2VA, FL2VA, and R2VA — seamless looping lives inside FL2VA via the "Loop Trim" bypass group
Vast.ai provisioning has been aligned with the current workflow set

🧠 QwenVL-Mod 2.10.2

The completed hackathon-only Livepeer Agent integration has been removed
Qwen Workflow Chat and the standard HF/GGUF prompt-enhancement pipeline remain fully available

⬆️ 2026/09/27 UPDATE ⬆️

🔄 Pure INT8 ConvRot — NVFP4 Removed

The MMH3 setup now uses pure INT8 ConvRot models. NVFP4 was removed after testing showed a noticeable quality loss. Speed comes from Turbo LoRAs, sparse attention, and the Triton backend.

FL2VA DiT: minimax_h3_fl2va_pruned_int8_convrot.safetensors (~21 GB) — Comfy-Org/MiniMax-H3
R2VA DiT: minimax_h3_ref2va_pruned_int8_convrot.safetensors (~21 GB) — Comfy-Org/MiniMax-H3
Uncensored H3 text encoder: qwen3vl_32b_heretic_minimax_h3_nvfp4.safetensors (~15 GB, full conditioning encoder) — Momoking
R2VA Native LoRA: minimax_h3_ref_lora_rank_256_bf16.safetensors (~2.6 GB; required for native R2VA on the plain ref2va checkpoint — NOT on Singularity, where it is fused) — Kijai
Singularity R2VA DiT ⭐: Minimax-h3_Singularity_ref2va_Pruned_v1.3_int8.safetensors (~21 GB, fused T2V/I2V/R2V/V2V) — WarmBloodAban/Minimax-h3_Singularity
10Eros: 10Eros_Max_h3_hybrid_beta5_int8.safetensors is a non-turbo checkpoint; use lightx2v_hybrid-4to8step-full-fusion_Turbo_pruned.safetensors for Turbo

⚙️ ComfyUI Launch Arguments

--highvram --disable-auto-launch --fast fp16_accumulation --enable-triton-backend --force-fp16

🔗 Model Chain — Order Matters

UNETLoader (INT8) → LoraLoader → ModelAttentionBackend (Comfy Kitchen Attention) → BlockSparseAttention (sol-attn) → MiniMaxH3SigmaShift → BasicGuider (CFG 1.0) → SamplerCustomAdvanced

If ModelAttentionBackend is missing from a workflow, add it manually between LoraLoader and BlockSparseAttention.

🎚️ Definitive Preset Table

PresetStepsCFGVideo ShiftAudio ShiftTauLoRA (strength 1.0)Native Base (FL2VA)28–321.06.03.01.00BypassedNative Turbo (FL2VA)81.012.04.01.30minimax_h3_fl2v_turbo_8step_v1.0_768p_comfyui_bf16R2VA Native25–301.04.03.01.00minimax_h3_ref_lora_rank_256_bf16R2VA Turbo81.06.03.01.30minimax_h3_ref2v_turbo_8step_v1.0_768p_comfyui_bf16 — euler + simpleR2VA Singularity ⭐201.04.03.01.00None — ref is fused in the checkpoint (res_multistep + simple)R2VA Singularity Turbo ⭐81.06.03.01.30minimax_h3_ref2v_turbo_8step_v1.0_768p_comfyui_bf16 — euler + simple10Eros Max Beta530–401.06.03.00.85–1.00Bypassed10Eros Turbo81.012.04.01.30lightx2v_hybrid-4to8step-full-fusion_Turbo_pruned

⬆️ 2026/09/23 UPDATE ⬆️

⚡ Native Block Sparse Attention — No Custom Node Required

All Turbo workflows now use the built-in ComfyUI BlockSparseAttention node instead of the custom ComfyUI-sol-attn pack (removed from GitHub upstream). The acceleration is identical — the native node runs the same sol-attn path:

Same speed as the custom node (verified A/B on Blackwell at 8 and 20 steps)
Method sol-attn, tau 1.3 — same threshold as before
sink_conditioning: exact_kv_and_rows protects text/audio/reference rows natively
FusedModulation and ChunkFeedForward dropped — no measurable speed difference
Nothing to install: works on stock ComfyUI ≥ 0.30, zero extra custom nodes

💬 Qwen Chat + Prompt Pipeline

New default chat model: GGUF: Qwen3.5-9B-Defiant-Fable-Uncnr-Heretic-NEO-MAX-Q8_0 — sharper uncensored routing
MiniMax output normalization (QwenVL-Mod ≥ 2.8.14): the enhancer now strips duplicated [Shot N] blocks before integrated_multimodal_description and auto-inserts the required I2VA reference-binding line (For the target video...) when missing — on both HF and GGUF backends
Preset-family routing fixed: asking "10 seconds" inside an FL2VA or R2VA workflow now keeps the same mode instead of falling back to the generic preset

📌 Requires QwenVL-Mod ≥ 2.8.14 (ComfyUI Manager → Update, or git pull).

⬆️ 2026/09/21 UPDATE ⬆️

🧠 Unified HF/GGUF QwenVL Nodes + Qwen 3.8 Default

All workflows now ship with the new unified QwenVL nodes — a single node with a backend dropdown instead of separate HF and GGUF nodes:

Backend dropdown: switch between HF transformers and GGUF llama.cpp in one click — no more rewiring the graph to change backend
GGUF default: workflows now default to GGUF: Qwen3.8-9B-heretic-uncensored.Q8_0.gguf — faster load, lower VRAM, uncensored out of the box
Existing HF workflows still work; the dropdown also accepts any GGUF you drop in models/LLM (auto-discovery, no JSON editing)

💬 Qwen Workflow Chat + Livepeer Agent Render

The pack's companion features kept evolving (see the QwenVL-Mod repo for the full changelog):

Qwen Chat sidebar: inspects the loaded workflow, edits widgets, picks presets and queues renders from natural language — understands the MiniMax H3 acceleration stack (Sol-Attn tau, Turbo LoRA, native steps) and switches configs like "10Eros", "Native" or "Turbo" on request
Livepeer Agent render node (QwenVL_LivepeerRender): optional node that renders images/videos through the Livepeer Agent network via MCP, with a source_video frame picker for i2v references — the chat drives it directly

📌 Requires QwenVL-Mod ≥ 2.8.0 (ComfyUI Manager → Update, or git pull). Older versions will show the new unified nodes as missing until updated.

⬆️ 2026/09/16 UPDATE ⬆️

🚫 Acceleration Stack Simplified — Spectrum + DiffAid Removed

Testing confirmed that Spectrum's block wrapper is code-incompatible with Sol fused blocks (crash: unexpected keyword argument 'attention', verified at 20 steps), and DiffAid showed no confirmed benefit. Both node packs have been removed from the stack entirely — they are no longer shipped or required.

Sol-Attn is now the only acceleration patch and stays active in ALL modes:
- 10Eros TURBO: tau 1.30
- Turbo LoRA: tau 1.30 (FL2VA) / 1.30 (R2VA)
- Native: tau 1.00
Sol-Fusion + Sol-FFN stay always ON (50 blocks / 52 MLPs)
The old "bypass Sol-Attn with Turbo" rule is superseded — high-tau Sol-Attn works fine on turbo
Workflows updated: no Spectrum/DiffAid nodes required anymore

💬 Workflow-Aware Qwen Chat

The Qwen chat assistant is now workflow-aware — it reads the loaded workflow's widgets and can act on them directly:

Edit widgets from chat: ask "change steps to 20" or "rewrite the prompt" and the assistant applies set_widget_value / bypass / queue_workflow actions on the real nodes — no manual clicking
Preset-aware: injects the system guide of the preset selected in the workflow (e.g. MiniMax H3 NSFW 5s/10s/15s), so answers match the actual preset rules
Duration switching: ask for a different clip length ("make it 10 seconds") and it picks the matching duration preset and updates the length/frame widgets automatically
Full prompt echo: when it rewrites a prompt widget, the complete new text is repeated verbatim in the reply — no hidden truncation
Choice buttons: for ambiguous requests (e.g. Turbo LoRA vs native) it asks first with clickable options instead of guessing
Knows the acceleration stack: understands the Sol-Attn modes and tau values, so "switch to native quality" sets sampler, steps, shift and tau correctly

Also in this update: Qwen3.5 native support — the new qwen3_5 architecture (hybrid linear/full attention) requires transformers>=5.2.0; older releases fail with "model type qwen3_5 not recognized".

⬆️ 2026/09/04 UPDATE ⬆️

🔧 Turbo LoRA Switch — larryvrh → lightx2v 8-step 768p

All Turbo workflows now use lightx2v Turbo LoRA 8-step 768p (Apache-2.0) instead of larryvrh v4-600:

FL2VA: minimax_h3_fl2v_turbo_8step_v1.0_768p_comfyui_bf16.safetensors (~1.96 GB)
Ref2VA: minimax_h3_ref2v_turbo_8step_v1.0_768p_comfyui_bf16.safetensors (~1.96 GB)
No custom node required — standard ComfyUI LoRA loader works
Removed Larryvrh/ComfyUI-MiniMax-H3-Turbo custom node
Trained at 1344×768 — matches our native resolution exactly

🚫 Acceleration Patch Rules with Turbo — Verified

Community testing informed the acceleration rules (see the 2026/09/16 update for the current stack):

Sol-Attn on Turbo LoRA: tau 1.30 is the verified value at 8 steps (per the preset table). On 10Eros turbo the same tau applies
Spectrum was removed from the stack — code-incompatible with Sol fused blocks
Rule: Sol-Attn stays on in every mode, only the tau changes

Turbo: Turbo LoRA ON · Sol-Attn tau 1.30 · euler + simple · 8 steps, shift 12/4 (FL2VA) o 6/4 (R2VA) Native: Turbo LoRA OFF · Sol-Attn tau 1.00 · res_multistep + simple · 28-32 steps, shift 6/3

🔄 10Eros-Max Switch — DmitryDB → cicalooo TURBO Hybrid Beta3

The 10Eros-Max model has been switched to cicalooo's ComfyUI-native INT8 ConvRot Turbo Hybrid Beta3:

Model: 10Eros_Max_h3_TURBO-hybrid_beta3_int8_convrot_skip_edges.safetensors (~22.5 GB)
TURBO fused in checkpoint — no separate t8star LoRA needed
Boundary blocks 0, 1, 48, 49 left in BF16 for stability
Removed DmitryDB 10Eros model + t8star compatibility LoRA

⚡ Comfy Kitchen Attention — Replaces SageAttention

All MMH3 Docker/provisioning now uses --use-ck-attention (Comfy Kitchen Attention) instead of --use-sage-attention:

Faster on Blackwell (RTX 5090 / PRO 6000)
Better detail preservation
No separate SageAttention package needed

⬆️ 2026/09/02 UPDATE ⬆️

The QwenVL node now has a camera_tag dropdown — no more typing [ORBIT] manually in the prompt. Select from 19 camera movements directly in the node UI:

CategoryTagsStatic[STATIC_CAMERA], [LOCKED_OFF]Slow zoom[SLOW_ZOOM_IN], [SLOW_ZOOM_OUT]Fast zoom[FAST_ZOOM_IN], [FAST_ZOOM_OUT]Pan[PAN_LEFT], [PAN_RIGHT]Tilt[TILT_UP], [TILT_DOWN]Dolly[DOLLY_IN], [DOLLY_OUT]Tracking[TRACKING_LEFT], [TRACKING_RIGHT]Crane[CRANE_UP], [CRANE_DOWN]Other[ORBIT], [HANDHELD], [ROLL]

How it works: when you select a tag, it's injected both at the start of the prompt and as a FINAL CAMERA DIRECTIVE at the end — so Qwen 9B actually respects it despite recency bias on long prompts. The tag also gets a short description so Qwen knows exactly what to write.

Subject stays alive: the directive explicitly tells Qwen that the camera tag controls ONLY the camera — the subject must still have natural, lively action (breathing, gestures, expression, body motion) throughout the clip. No more "statue during orbit" problem.

Manual tags still work: if you leave the dropdown on None and type [ORBIT] in your prompt, it's detected and injected automatically as a fallback.

Available on all three QwenVL nodes: AILab_QwenVL, AILab_QwenVL_Advanced, and AILab_QwenVL_PromptEnhancer.

🔄 FL2VA Loop Merged into FL2VA — One Workflow, Bypass Group

The loop trim logic now lives inside the main FL2VA workflow, wrapped in a "Loop Trim" group that can be toggled via the rgthree Fast Groups Bypasser node:

Loop mode (trim active): the ImageFromBatch + ComfyMathExpression nodes trim the frozen tail (~5 frames) for seamless looping
Non-loop mode (trim bypassed): toggle the group off in the Bypasser → VAEDecode passes directly to RIFE/upscale, full frames preserved

No more switching between two workflows — just toggle the group.

🧹 PromptEnhancer Cleanup

Removed the redundant custom_system_prompt input — enhancement_style (presets) + prompt_text (user input) cover all use cases
Removed CUSTOM_ONLY_STYLE ("✍️ Custom Only (no preset)") — no longer needed
Added camera_tag dropdown (same as main QwenVL nodes)

📦 Workflow Count

With the loop merged into FL2VA, the pack ships 4 workflows (T2VA, I2VA, FL2VA, R2VA) plus the combined ALL-WFs zip. All workflows updated with the new camera_tag input.

⬆️ 2026/08/31 UPDATE ⬆️

🎲 Wildcards (T2VA Workflow)

The T2VA Turbo workflow includes a WildcardProcessor node that injects randomized prompt fragments from the PMP's Prompt Engine (__pmp/prmpt/*) wildcard library. Each generation picks a random entry from each wildcard file, so the same seed produces different results across runs unless you pin the seed.

Wildcards Used in the T2VA Workflow

WildcardCategoryWhat it randomizes__pmp/prmpt/vidstyle__Video styleCinematic / anime / vintage film / 3D CG / claymation / watercolor / fantasy etc.__pmp/prmpt/imgcmp/shots__Camera shotClose-up / wide / medium / dolly / crane / orbit / handheld etc.__pmp/prmpt/lctns/rndmlctns__LocationRandom setting (bedroom / beach / studio / alley / forest / rooftop etc.)__pmp/prmpt/light/rndmlight__LightingKey light direction, color temperature, soft/hard, ambient mood__pmp/prmpt/char/favchar__CharacterRandom character archetype (age, body type, hair, ethnicity)__pmp/prmpt/clths/rndmclth__ClothingRandom outfit / garment description__pmp/prmpt/clths/nudty_brsts__NSFWBreast / nudity descriptors (NSFW preset)__pmp/actsolo/brstplng__NSFW actionSolo breast-play action verbs (NSFW preset)

How It Works

The WildcardProcessor node sits before the Qwen3.5 prompt enhancer
At queue time, each __wildcard__ token is replaced with a random line from the corresponding .txt file inside ComfyUI/custom_nodes/ComfyUI-Wildcards/wildcards/pmp/prmpt/
The expanded text is passed to Qwen3.5, which converts it into the official MiniMax H3 prompt format
Different seed = different wildcard picks — use a fixed seed if you want reproducible results

Customizing Wildcards

Edit existing: open the .txt files under ComfyUI/custom_nodes/ComfyUI-Wildcards/wildcards/pmp/prmpt/ and add/remove lines (one entry per line)
Add your own: create a new .txt file, e.g. pmp/prmpt/mytags.txt, then reference it as __pmp/prmpt/mytags__
Remove a wildcard: delete the __...__ token from the WildcardProcessor text field in the workflow
Disable randomization: replace the __wildcard__ token with a fixed string

Required Custom Node

ComfyUI-TagForge (includes the WildcardProcessor node and the __pmp/prmpt/* wildcard set) — huchukato/ComfyUI-TagForge

The wildcard files ship with the custom node. If a wildcard resolves to empty, the custom node is missing or the wildcard folder is not installed.

🔄 Sampler Change — MiniMaxH3TurboSampler → KSamplerSelect + MiniMaxH3SigmaShift

All Turbo workflows (T2VA, I2VA, FL2VA, FL2VA-Loop, R2VA) have been updated to use ComfyUI core nodes instead of the custom MiniMaxH3TurboSampler:

Removed: MiniMaxH3TurboSampler (custom node from Larryvrh/ComfyUI-MiniMax-H3-Turbo)
Added: KSamplerSelect (sampler: euler) + MiniMaxH3SigmaShift (shift_video=12, shift_audio=3) — both ComfyUI core nodes, no custom node required
Scheduler: simple (unchanged)

Why?

On ComfyUI v0.35.0+ with native ModelSamplingAV, the custom MiniMaxH3TurboSampler internally delegates to stock euler anyway — the custom node is redundant
Using core nodes means the same workflow works with both:
- Standard model (minimax_h3_fl2va_pruned_int8_convrot) + Turbo LoRA minimax_h3_turbo_v4_step600_ema
- 10Eros-Max (10Eros_Max_H3_FL2VA-INT8-ConvRot-HQ) + T8 compatibility LoRA minimax_h3_fl2v_turbo_8step_v1.0_10ErosMax_beta1_pruned_compat_v001_T8
Just swap LoadDiffusionModel and LoraLoaderBypassModelOnly — the sampler path stays the same

10Eros-Max (Optional — Experimental)

Model: 10Eros_Max_H3_FL2VA-INT8-ConvRot-HQ.safetensors (~23.5 GB) — DmitryDB/MiniMax-H3-10Eros-Max-Quants
LoRA: minimax_h3_fl2v_turbo_8step_v1.0_10ErosMax_beta1_pruned_compat_v001_T8.safetensors (~1.96 GB) — t8star/minimax_h3_turbo_4step_10ErosMax_test4_pruned_curveproj1025_T8
Sampler: euler + MiniMaxH3SigmaShift (shift 12/4) + simple scheduler — same as standard model
⚠️ NVFP4 rimosso — perdita di qualità; usa solo INT8 ConvRot (10Eros_Max_h3_hybrid_beta5_int8 + LoRA lightx2v_hybrid-4to8step)
⚠️ The T8 LoRA is checkpoint-specific — only works with the exact 10Eros pruned model (SHA-256: f82cc3f723b080e7ae94a7c98f95aa989e387618d0bdc940133dfbd9f432c062)

⬆️ 2026/08/27 UPDATE ⬆️

Pure INT8 ConvRot — Default

Default diffusion models: pure INT8 ConvRot (minimax_h3_fl2va_pruned_int8_convrot / minimax_h3_ref2va_pruned_int8_convrot) from Comfy-Org/MiniMax-H3 — best quality; NVFP4 was removed due to visible quality loss
Uncensored text encoder: NVFP4 (qwen3vl_32b_heretic_minimax_h3_nvfp4) from Momoking/Qwen3-VL-32B-Heretic-MiniMax-H3-NVFP4 — full H3 conditioning, ~15 GB
Speed comes from turbo LoRA (lightx2v 8-step), sol-attn sparse attention, and the --enable-triton-backend comfy-kitchen backend — works on any modern NVIDIA

SOL-ATTN Integration

All Turbo workflows include SOL-ATTN nodes (sparse attention + fused modulation + chunked FFN)
Tau per mode: 1.30 su turbo (FL2VA/R2VA/10Eros), 1.00 su native — vedi tabella preset in testa
Spectrum and DiffAid were removed from the stack — Sol-Attn is the only acceleration patch

Turbo Step Standardization

All Turbo workflows standardized to 8 steps (was 6 for T2VA/R2VA)
Consistent minimax_h3_fl2v_turbo_8step_v1.0_768p_comfyui_bf16 LoRA across all workflows

TensorRT Batch Size

RIFE and Upscaler TRT expose separate loader and runner batch_size parameters
Verified stable configuration:
- RIFE: loader 1, runner 1
- Upscaler: loader 2, runner 2
The loader compiles the TensorRT engine profile; the runner controls frames sent per infer() call (runner ≤ loader)
RIFE batch values above 1 can build but currently fail during interpolation; keep RIFE at 1/1
Upscaler 2/2 is verified on RTX PRO 6000 Blackwell; batch 4 fails to build with TensorRT 10.15 on sm_120
Changing the loader batch size requires a different engine; delete incompatible cached TRT engines before rebuilding

Upscaler: Auto-detect Scale Factor

Removed the scale dropdown (2x/4x) from the Upscaler runner node — it was redundant and error-prone
The loader now auto-detects the upscale factor from the model name (2x* → 2, 4x* → 4, x2plus → 2, x4plus → 4)
The factor is passed to the runner via the engine object — no more mismatch between model and scale setting
Requires ComfyUI-Upscaler-TensorRT-Auto updated to latest version

⚠️ Requirements — Read First!

GPU & VRAM

🟢 Recommended template configuration — RTX 5090 (32 GB) / RTX PRO 6000 (48 GB) → INT8 ConvRot diffusion + INT8 Heretic text encoder
🟡 Low-VRAM alternative — RTX 4090 / 3090 (24 GB) → INT8 ConvRot + offload
🟠 Lower-VRAM alternative — 12–16 GB → INT4 + aggressive offload; slow and not recommended for production

12 GB GPUs (e.g. RTX 3060 12GB): Technically possible with INT4 models + aggressive offloading, but very slow. You need 32 GB+ system RAM and a fast NVMe SSD. Not recommended for production use.

Model Quantization Options

BF16 (full) — Diffusion ~42 GB + Text encoder ~65 GB = ~110 GB total → Comfy-Org/MiniMax-H3
INT8 (pruned) — Diffusion ~21 GB + Text encoder ~24.5 GB = ~50 GB total → Comfy-Org/MiniMax-H3
INT4 (pruned) — Diffusion ~11 GB + Text encoder ~15 GB = ~24.5 GB total → Merserk/MiniMax-H3-INT4-ConvRot
Pure INT8 ConvRot (recommended — verified) — Diffusion ~21 GB + uncensored INT8 Heretic text encoder ~26.4 GB = ~47 GB active model set → Comfy-Org/MiniMax-H3 + ethanfel/Qwen3-VL-32B-Ultra-Heretic-H3-ComfyUI-INT8-ConvRot. Best quality on any modern NVIDIA

Software

ComfyUI: current upstream master (required for MiniMax H3 native support)
Python: 3.10+
CUDA: 12.8+ (13.0 recommended)
Storage: allow at least 130 GB for the complete provisioned package (~94 GB of models plus engines, workflows and outputs)

Qwen3.5 Prompt Enhancer

GGUF: Q4_K_S or Q5_K_S quantization for 4B/9B models
HF: Qwen3.5-9B-Defiant-Fable-Heretic (~18 GB) or Qwen3.5-4B-heretic-v2 (~8 GB)

⚡ MiniMax-H3 Turbo LoRA (Optional — Faster & Sharper)

A distilled 8-step LoRA for MiniMax-H3 that replaces the default 28-32 step sampling. All Turbo workflows include the native BlockSparseAttention (sol-attn) node — tau 1.30 at 8 steps.

Sparse attention: built into ComfyUI — the native BlockSparseAttention node (method sol-attn), no custom node required
Recommended LoRA: minimax_h3_fl2v_turbo_8step_v1.0_768p_comfyui_bf16.safetensors (~1.96 GB)
Download: lightx2v/Minimax-h3-Turbo
Install: place the .safetensors in ComfyUI/models/loras/
Usage: 8 steps with scheduler simple, sampler euler + MiniMaxH3SigmaShift (shift 12/4 FL2VA, 6/4 R2VA). No custom sampler node required — standard LoRA loader works.

Works with all tasks: T2VA, I2VA, FL2VA and R2VA.

⚙️ Model Configuration Cheat Sheet

All Turbo workflows ship with bypass groups for Sol-Attn and Turbo LoRA. Toggle them via the rgthree Fast Groups Bypasser node depending on which model you load.

📊 Configuration Matrix

Setting10Eros Turbolightx2v Turbo LoRANativeDiffusion model10Eros_Max_h3_hybrid_beta5_int8minimax_h3_fl2va_pruned_int8_convrotminimax_h3_fl2va_pruned_int8_convrotTurbo LoRA✅ ON lightx2v_hybrid-4to8step✅ ON (strength 1.0)❌ OFFSteps8828-32Samplereulereulerres_multistepSchedulersimplesimplesimpleVideo shift12126Audio shift443Sol-Attn✅ ON (tau 1.30)✅ ON (tau 1.30)✅ ON (tau 1.00)CK Attention✅ ON (--enable-triton-backend)✅ ON✅ ON

🔧 How to Switch Models in the Workflow

LoadDiffusionModel — swap the .safetensors file
LoraLoaderBypassModelOnly — toggle bypass:
- 10Eros / Native → bypassed (LoRA off)
- Turbo LoRA → active (strength 1.0)
Sol-Attn node — set tau (or bypass the group via Fast Groups Bypasser):
- 10Eros → tau 1.30
- Turbo LoRA → tau 1.5-2.0 or bypassed
- Native → tau 1.0
KSamplerSelect — change sampler (euler for Turbo/10Eros, res_multistep for Native)
MiniMaxH3SigmaShift — shift 12/4 (FL2VA turbo), 6/4 (R2VA turbo), 6/3 (10Eros), 4/3 (R2VA native), 6/3 (native FL2VA)
Sampler steps — 8 turbo, 28-32 native, 30-40 10Eros

🚀 10Eros-Max TURBO Hybrid (Recommended for Turbo)

Model: 10Eros_Max_h3_hybrid_beta5_int8.safetensors (~20 GB) — TenStrip/10Eros-Max
Non-turbo checkpoint: per la modalità turbo applicare il LoRA lightx2v_hybrid-4to8step-full-fusion_Turbo_pruned (strength 1.0)
Sol-Attn ON — tau 1.30 per il preset 10Eros turbo
10Eros Beta5 è non-turbo: per la modalità turbo applicare il LoRA lightx2v_hybrid-4to8step (strength 1.0)

⚡ lightx2v Turbo LoRA (Standard Turbo)

LoRA: minimax_h3_fl2v_turbo_8step_v1.0_768p_comfyui_bf16.safetensors (~1.96 GB) — lightx2v/Minimax-h3-Turbo
Ref2VA LoRA: minimax_h3_ref2v_turbo_8step_v1.0_768p_comfyui_bf16.safetensors (~1.96 GB) — same source
Trained at 1344×768 — matches native resolution
Sol-Attn: tau 1.5-2.0, or OFF for max safety (tau 1.0 causes fallbacks at 8 steps)
No custom node required — standard LoRA loader works

🎬 Native (Maximum Quality)

Model: standard pure INT8 ConvRot (minimax_h3_fl2va_pruned_int8_convrot)
No Turbo LoRA — full 28-32 step sampling
Sol-Attn ON (tau 1.00) — safe on the native trajectory
Sampler: res_multistep (not euler)
Shift: 6/3 (FL2VA native), 4/3 (R2VA native con ref_lora_rank_256)
Slower but highest visual quality

🌟 What is ComfyUI-QwenVL-Mod?

A powerful enhanced vision-language node for ComfyUI that combines Qwen3.5 models with MiniMax H3 video generation workflows. Features multilingual support, visual style detection, native stereo audio, and NSFW capabilities for professional AI content creation.

Think: "Your all-in-one solution for intelligent prompt enhancement and video+audio generation with MiniMax H3!"

🎬 Key Features

🚀 MiniMax H3 Video+Audio Generation

T2VA (Text-to-Video+Audio): Generate video with native stereo audio from text
I2VA (Image-to-Video+Audio): Animate a first-frame image with audio
FL2VA (First-Last-Frame): Generate the transition between two keyframes — Qwen3.5 sees both frames
R2VA (Reference-to-Video): Lock character identity, style, motion, or voice using reference images

🧠 Qwen3.5 Auto-Prompting

Multilingual: Write your prompt in any language — Qwen3.5 translates and converts it
Auto-format: Generates the official MiniMax H3 prompt format (3-field for base, 6-field for R2VA)
Multi-reference: Qwen3.5 analyzes the two images connected through image + image2; additional MiniMax references can be described with [P3] / [P4] tags
Visual style detection: 12+ artistic styles (photorealistic, cinematic, anime, 3D CG, claymation, vintage film, watercolor, fantasy, etc.)
Smart caching: Performance optimization with Fixed Seed Mode
GGUF backend: Efficient local model inference with quantization support
Qwen3.5 support: Thinking mode disabled via /no_think for fast prompt generation

🔊 Native Stereo Audio

No separate audio node needed — MiniMax H3 generates video and audio jointly in a single forward pass
Voice, sound effects, and music modeled together, not layered on afterward
Describe sounds in your prompt and the model generates them natively

🎨 NSFW Support

Comprehensive content generation without restrictions
9 dedicated NSFW presets (3 base 🎬 + 3 R2VA 🎞️ + 3 FL2VA 🔄) with explicit diegetic soundscape
Natural progression, style adaptation, consistent characters

📦 What's Included — 5 Turbo Workflows

All workflows are pre-wired with lightx2v Turbo LoRA at 8 steps + SOL-ATTN (tau per mode — see the configuration matrix).

📥 Download

FileContentsLinkMiniMaxH3-Turbo-Qwen3.5-ALL-WFs.zipAll 5 workflows (T2VA + I2VA + FL2VA + FL2VA-Loop + R2VA)DownloadMiniMaxH3-Turbo-T2VA-Qwen3.5.zipT2VA onlyDownloadMiniMaxH3-Turbo-I2VA-Qwen3.5.zipI2VA onlyDownloadMiniMaxH3-Turbo-FL2VA-Qwen3.5.zipFL2VA only (includes bypassable loop trim)DownloadMiniMaxH3-Turbo-FL2VA-Loop-Qwen3.5.zipFL2VA Loop only (seamless looping)DownloadMiniMaxH3-Turbo-R2VA-Qwen3.5.zipR2VA onlyDownload

Individual .json files also available in workflows/minimax/.

Workflows

⚡ T2VA Turbo — MiniMaxH3-Turbo-T2VA-Qwen3.5.json — text only — Text-to-video+audio. Simplest workflow.
⚡ I2VA Turbo — MiniMaxH3-Turbo-I2VA-Qwen3.5.json — text + first-frame image (image) — Image-to-video. First-frame animation with audio.
⚡ FL2VA Turbo — MiniMaxH3-Turbo-FL2VA-Qwen3.5.json — text + first-frame (image) + last-frame (image2) — First-Last-Frame to video. Includes TensorRT upscale + RIFE frame interpolation for 48 fps output. Loop trim is built in — toggle the "Loop Trim" group via the Fast Groups Bypasser node for seamless loops.
⚡ R2VA Turbo — MiniMaxH3-Turbo-R2VA-Qwen3.5.json — text + reference images (image + image2) — Reference-to-video. Lock identity, style, motion, camera, or voice using up to 9 ref images.

FL2VA and R2VA include TensorRT upscaling and RIFE frame interpolation for 48 fps high-resolution output.

🖼️ Multi-Reference Input (image2)

The QwenVL-Mod node has two image inputs:

T2VA: no images needed
I2VA: image = first frame
FL2VA: image = first frame, image2 = last frame, frame_count = 1
R2VA: image = primary reference, image2 = additional references (batch, up to 9), frame_count = 1–9

Qwen3.5 sees all connected images as individual images (not as a video sequence), enabling proper multi-reference analysis for FL2VA and R2VA.

🎯 QwenVL-Mod NSFW Presets (9 total)

The workflows include built-in NSFW presets for the Qwen3.5 prompt enhancer:

🎬 Base Presets (T2VA / I2VA)

🎬 MiniMax H3 NSFW (5s) — 5 seconds — 3 fields: integrated_multimodal_description + overall_soundscape + non_diegetic_music
🎬 MiniMax H3 NSFW (10s) — 10 seconds — Same format
🎬 MiniMax H3 NSFW (15s) — 15 seconds — Same format

🔄 FL2VA Presets (First-Last-Frame)

🔄 MiniMax H3 NSFW FL2VA (5s) — 5 seconds — 3 fields, transition-focused (describes the path between frames)
🔄 MiniMax H3 NSFW FL2VA (10s) — 10 seconds — Same format
🔄 MiniMax H3 NSFW FL2VA (15s) — 15 seconds — Same format

🎞️ R2VA Presets (Reference)

🎞️ MiniMax H3 NSFW R2VA (5s) — 5 seconds — 6 fields: subject_definitions + summary + retention_analysis + detailed_description + overall_soundscape + non_diegetic_music
🎞️ MiniMax H3 NSFW R2VA (10s) — 10 seconds — Same format
🎞️ MiniMax H3 NSFW R2VA (15s) — 15 seconds — Same format

What the presets produce

🎬 Base: [Shot 1] with style + initial composition, camera vocabulary, speaker IDs, diegetic soundscape
🔄 FL2VA: Describes the transition path between first and last frames (not the scene — images fix the scene). Favors single continuous shot.
🎞️ R2VA: 6-section format with <Subject N>, <Picture N>, <Video N>, <Audio N> labels, retention markers (fully_preserved, partially_preserved, etc.), task-type summary
All presets: smooth, continuous camera motion (no abrupt or stepped changes), explicit diegetic soundscape, optional non-diegetic music (defaults to N/A)
All presets: support camera control tags ([STATIC_CAMERA], [SLOW_ZOOM_IN], [SLOW_ZOOM_OUT], [ORBIT], [HANDHELD]) — see Camera Control Tags below
FL2VA presets: automatic loop mode when first and last frame are the same image — see Loop Mode below

SFW presets are also available. Edit the preset dropdown in the QwenVL node to switch.

🎮 Usage Examples

Basic Text-to-Video (T2VA)

Load MiniMaxH3-Turbo-T2VA-Qwen3.5.json
Write your prompt in any language
Select preset 🎬 MiniMax H3 NSFW (5s/10s/15s)
Generate video with native audio

Image-to-Video (I2VA)

Load MiniMaxH3-Turbo-I2VA-Qwen3.5.json
Upload your first-frame image to image
Select preset 🎬 MiniMax H3 NSFW (5s/10s/15s)
Write what happens next (in any language)
Generate animated video with audio

First-Last-Frame (FL2VA)

Load MiniMaxH3-Turbo-FL2VA-Qwen3.5.json
Upload first-frame to image, last-frame to image2, set frame_count=1
Select preset 🔄 MiniMax H3 NSFW FL2VA (5s/10s/15s)
Describe the transition between the two frames
Generate the interpolated video at 48 fps with TensorRT upscale + RIFE

Reference-to-Video (R2VA)

Load MiniMaxH3-Turbo-R2VA-Qwen3.5.json
Upload primary reference to image, additional references to image2 (batch), set frame_count to match
Select preset 🎞️ MiniMax H3 NSFW R2VA (5s/10s/15s)
Reference them by tag in your prompt: <Picture 1>, <Picture 2>, etc.
Generate video with locked identity/style

🔧 Technical Specifications

⚡ Performance

Output: 768p, 24 fps (native), up to ~15 seconds
Audio: Native stereo, generated jointly with video
Upscale: TensorRT RealESRGAN x4 (FL2VA + R2VA workflows)
Frame interpolation: RIFE v4.25 → 48 fps (FL2VA + R2VA workflows)
Comfy Kitchen Attention (--enable-triton-backend): triton backend, FP16 accumulation
Smart caching: Reuse prompts with same inputs, Fixed Seed Mode for text-only caching

🎨 Model Support

Qwen3.5: 4B / 9B / 27B (uncensored, heretic, unsloth) — thinking mode disabled
Qwen3.8: latest-generation models (GGUF + HF)
HF Models: Josiefed, official, Heretic-Stable variants
Quantization: Q4_K_S, Q5_K_S, FP16, INT8

🌐 Multilingual Capabilities

Input languages: Any language supported
Auto-translation: Automatic translation to optimized English
Style detection: Works with multilingual prompts
Cultural adaptation: Context-aware prompt enhancement

📦 Installation

Quick Install

Download: ComfyUI-QwenVL-Mod (latest version)
Extract to ComfyUI/custom_nodes/ComfyUI-QwenVL-Mod
Install requirements: pip install -r requirements.txt
Restart ComfyUI
Load included workflows from minimax/ folder

Custom Nodes Required

ComfyUI-QwenVL-Mod — All workflows (Qwen3.5 prompt enhancer) — huchukato/ComfyUI-QwenVL-Mod
Block Sparse Attention — Turbo workflows use the native ComfyUI node (built-in, no install needed)
ComfyUI-RIFE-TensorRT-Auto — FL2VA, R2VA (frame interpolation) — huchukato/ComfyUI-RIFE-TensorRT-Auto
ComfyUI-Upscaler-TensorRT-Auto — FL2VA, R2VA (upscaling) — huchukato/ComfyUI-Upscaler-TensorRT-Auto
ComfyUI-VideoHelperSuite — FL2VA, R2VA (VHS_VideoCombine) — Kosinkadink/ComfyUI-VideoHelperSuite
ComfyUI-Easy-Use — FL2VA, R2VA (easy showAnything) — yolain/ComfyUI-Easy-Use
ComfyUI-PerfectVideoResolution — All workflows (resolution calculator) — huchukato/ComfyUI-PerfectVideoResolution

Note: ComfyMathExpression is built into ComfyUI core (v0.24.1+) — no custom node needed.

Models Required

T2VA / I2VA / FL2VA use the INT8 FL2VA model:

models/vae/ → minimax_h3_video_vae_fp16.safetensors (~5 GB)
models/vae/ → minimax_h3_audio_vae_fp32.safetensors (~0.6 GB)
models/diffusion_models/ → minimax_h3_fl2va_pruned_int8_convrot.safetensors (~20 GB) — Comfy-Org/MiniMax-H3
models/text_encoders/ → qwen3vl_32b_heretic_minimax_h3_nvfp4.safetensors (~15 GB) — Momoking/Qwen3-VL-32B-Heretic-MiniMax-H3-NVFP4

R2VA (ref2va) uses the INT8 Ref2VA model:

models/diffusion_models/ → minimax_h3_ref2va_pruned_int8_convrot.safetensors (~20 GB) — Comfy-Org/MiniMax-H3

INT8 ConvRot runs on any modern NVIDIA GPU — no Blackwell requirement.

10Eros-Max INT8 (optional/experimental)

models/diffusion_models/ → 10Eros_Max_H3_FL2VA-INT8-ConvRot-HQ.safetensors (~23.5 GB)
Pair it only with minimax_h3_fl2v_turbo_8step_v1.0_10ErosMax_beta1_pruned_compat_v001_T8.safetensors (~1.96 GB)
Switch both diffusion model and matching LoRA together; do not mix the standard and 10Eros LoRAs
The standard MiniMax H3 pure INT8 ConvRot is the verified default. 10Eros-Max (10Eros_Max_h3_hybrid_beta5_int8) remains experimental — non-turbo checkpoint, apply the lightx2v_hybrid-4to8step LoRA for turbo mode

Turbo LoRA (standard — all tasks)

models/loras/ → minimax_h3_fl2v_turbo_8step_v1.0_768p_comfyui_bf16.safetensors (~1.96 GB)
models/loras/ → minimax_h3_ref2v_turbo_8step_v1.0_768p_comfyui_bf16.safetensors (~1.96 GB)
No custom node required — standard LoRA loader works

INT4 alternative (for 12-16 GB GPUs): Merserk/MiniMax-H3-INT4-ConvRot

Qwen Prompt Enhancer

models/LLM/ → Qwen3.5-9B-Defiant-Fable-Heretic or Qwen3.5-4B-heretic-v2 (GGUF or HF)

TensorRT Engines (FL2VA + R2VA only)

models/upscale_models/ → RealESRGAN_x4 (TensorRT engine)
models/rife/ → rife425_ensemble_False_scale_1_sim (TensorRT engine, ONNX auto-downloaded from HF)

TensorRT engines must be built for your specific GPU. See ComfyUI-RIFE-TensorRT-Auto and ComfyUI-Upscaler-TensorRT-Auto for build instructions.

Download Links

VAE: video_vae_fp16 · audio_vae_fp32
Diffusion (fl2va, INT8 ConvRot): minimax_h3_fl2va_pruned_int8_convrot.safetensors
Diffusion (ref2va, INT8 ConvRot): minimax_h3_ref2va_pruned_int8_convrot.safetensors
Diffusion (fl2va, INT8 — Comfy-Org path): minimax_h3_fl2va_pruned_int8_convrot.safetensors
Diffusion (ref2va, INT8 — Comfy-Org path): minimax_h3_ref2va_pruned_int8_convrot.safetensors
Text encoder (uncensored, NVFP4): qwen3vl_32b_heretic_minimax_h3_nvfp4.safetensors
Text encoder (official, INT8 alternative): qwen3vl_32b_minimax_h3_int8_convrot.safetensors
Turbo LoRA (standard, FL2VA): minimax_h3_fl2v_turbo_8step_v1.0_768p_comfyui_bf16.safetensors
Turbo LoRA (standard, Ref2VA): minimax_h3_ref2v_turbo_8step_v1.0_768p_comfyui_bf16.safetensors
10Eros-Max diffusion (optional/experimental, INT8): 10Eros_Max_h3_hybrid_beta5_int8.safetensors + LoRA lightx2v_hybrid-4to8step-full-fusion_Turbo_pruned
INT4 models: Merserk/MiniMax-H3-INT4-ConvRot

🎬 MiniMax H3 Prompting Notes

How to Write Your Prompt

Describe the scene naturally. Be clear about the concepts below — Qwen3.5 handles the rest:

🎨 Visual style (put it first): photorealistic, cinematic, anime, 3D CG, claymation, vintage film, watercolor, fantasy
👥 Subjects: number, gender, appearance, clothing, position, expression
🏃 Action / motion: what happens, speed, interaction
🎥 Camera: dolly, pan, zoom, static, handheld, crane, orbit — smooth and continuous (no abrupt changes)
🌍 Environment: setting, lighting, atmosphere, time of day
🔊 Audio (important!): dialogue, breaths, moans, skin contact, ambient sounds, music

🔄 FL2VA: Describe the transition between frames, not the scene (images fix the scene) 🎞️ R2VA: Reference inputs by tag: <Picture 1>, <Picture 2>, <Video 1>, <Audio 1>

Resolution Guidance

MiniMax H3 native canvas: 768 px short edge, long edge capped at 1344 px, multiples of 32.

📱 Portrait: 768×1344 · 896×1152 · 960×1280
⬛ Square: 1024×1024
🖥️ Landscape: 1344×768 · 1152×896 · 1280×960

⚠️ Match the aspect ratio to your input image! Forcing 16:9 on a portrait image will squash it.
⚠️ Avoid direct 1080p. Generate at native resolution, then upscale with TensorRT nodes (FL2VA + R2VA workflows).

Duration

Choose a preset: 5s / 10s / 15s. The Math Expression node snaps the frame count to the model's 17-frame-per-block grid (17k+5 at 24 fps).

🎥 Camera Control Tags

All MiniMax H3 NSFW presets support camera control via the camera_tag dropdown on the QwenVL node — no need to type tags manually. Select from 19 camera movements:

TagEffect[STATIC_CAMERA] / [LOCKED_OFF]Camera completely static — no zoom, pan, orbit, or any motion[SLOW_ZOOM_IN]Slow continuous push-in (dolly toward subject)[SLOW_ZOOM_OUT]Slow continuous pull-back (dolly away from subject)[FAST_ZOOM_IN]Fast aggressive push-in, dramatic[FAST_ZOOM_OUT]Fast pull-back, reveal context[PAN_LEFT]Smooth horizontal pan from right to left[PAN_RIGHT]Smooth horizontal pan from left to right[TILT_UP]Smooth vertical tilt from bottom to top, revealing the subject[TILT_DOWN]Smooth vertical tilt from top to bottom[DOLLY_IN]Physical dolly movement toward the subject (parallax, not optical zoom)[DOLLY_OUT]Physical dolly movement away from the subject (parallax)[TRACKING_LEFT]Lateral tracking shot moving left, subject stays in frame[TRACKING_RIGHT]Lateral tracking shot moving right, subject stays in frame[CRANE_UP]Crane/jib movement rising upward, revealing the scene from above[CRANE_DOWN]Crane/jib movement descending toward the subject[ORBIT]Smooth 360-degree orbit around the subject[HANDHELD]Subtle handheld sway with natural micro-movements[ROLL]Slow camera roll (rotation around the lens axis)

How it works: the selected tag is injected at the start of the prompt AND as a FINAL CAMERA DIRECTIVE at the end, so Qwen respects it despite recency bias on long prompts. The subject stays alive and active — the tag controls only the camera.

If the dropdown is set to None, Qwen3.5 chooses a natural camera movement automatically. You can also type tags manually in your prompt as a fallback.

Example:

Dropdown: [ORBIT]
Prompt: she continues a slow rhythmic motion, breathing steadily

🔄 Loop Mode (FL2VA Only)

The FL2VA presets include automatic loop mode detection. When you load the same image as both first frame (image) and last frame (image2), the preset detects the identical endpoints and generates a seamless cyclic action:

The motion starts immediately from frame 0 (no wind-up or preparation)
The action continues at a steady rhythm for the entire duration (no early freeze)
The final state matches the first frame exactly (pose, framing, expression)
For repetitive actions (oral, stroking, thrusting, grinding): the rhythm continues without interruption, with natural variations in pace, depth, and angle
The word "loop" or "repeat" is never used in the generated prompt — the cyclicity is implicit
Camera motion in loop mode uses continuous circular or oscillating movements that return to the starting position (combine with [STATIC_CAMERA] if you want a locked-off loop)

To use loop mode:

Load MiniMaxH3-Turbo-FL2VA-Qwen3.5.json (the main FL2VA workflow — loop is built in)
Upload the same image to both image (first frame) and image2 (last frame)
Select preset 🔄 MiniMax H3 NSFW FL2VA (5s/10s/15s)
Describe the action — the preset handles the cyclic structure automatically
(Optional) Set camera_tag to [STATIC_CAMERA] if you want no camera movement

Loop Trim Bypass Group: the FL2VA workflow includes a "Loop Trim" group (wrapped around ImageFromBatch + ComfyMathExpression) controlled by the rgthree Fast Groups Bypasser node:

Group ACTIVE (default) → trim removes the frozen tail (~5 frames) for seamless looping
Group BYPASSED → full frames preserved, VAEDecode passes directly to RIFE/upscale (non-loop use)

✂️ Automatic trim: The trim removes the last 5 frames (0.2s at 24fps) — the frozen tail that MiniMax H3 adds at the end of FL2VA generation. The ComfyMathExpression node calculates the trim length automatically from the duration:
5s → 119 (124 - 5)
10s → 238 (243 - 5)
15s → 357 (362 - 5)

⚠️ Limitations: The automatic trim removes the frozen tail but minor discontinuity at the cut point may still occur due to velocity or camera phase differences. For a pixel-perfect loop, crossfade the last 0.5s with the first 0.5s in post-production.

🎲 Wildcards (T2VA Workflow)

The T2VA Turbo workflow includes a WildcardProcessor node that injects randomized prompt fragments from the PMP's Prompt Engine (__pmp/prmpt/*) wildcard library. Each generation picks a random entry from each wildcard file, so the same seed produces different results across runs unless you pin the seed.

Wildcards Used in the T2VA Workflow

WildcardCategoryWhat it randomizes__pmp/prmpt/vidstyle__Video styleCinematic / anime / vintage film / 3D CG / claymation / watercolor / fantasy etc.__pmp/prmpt/imgcmp/shots__Camera shotClose-up / wide / medium / dolly / crane / orbit / handheld etc.__pmp/prmpt/lctns/rndmlctns__LocationRandom setting (bedroom / beach / studio / alley / forest / rooftop etc.)__pmp/prmpt/light/rndmlight__LightingKey light direction, color temperature, soft/hard, ambient mood__pmp/prmpt/char/favchar__CharacterRandom character archetype (age, body type, hair, ethnicity)__pmp/prmpt/clths/rndmclth__ClothingRandom outfit / garment description__pmp/prmpt/clths/nudty_brsts__NSFWBreast / nudity descriptors (NSFW preset)__pmp/actsolo/brstplng__NSFW actionSolo breast-play action verbs (NSFW preset)

How It Works

The WildcardProcessor node sits before the Qwen3.5 prompt enhancer
At queue time, each __wildcard__ token is replaced with a random line from the corresponding .txt file inside ComfyUI/custom_nodes/ComfyUI-Wildcards/wildcards/pmp/prmpt/
The expanded text is passed to Qwen3.5, which converts it into the official MiniMax H3 prompt format
Different seed = different wildcard picks — use a fixed seed if you want reproducible results

Customizing Wildcards

Edit existing: open the .txt files under ComfyUI/custom_nodes/ComfyUI-Wildcards/wildcards/pmp/prmpt/ and add/remove lines (one entry per line)
Add your own: create a new .txt file, e.g. pmp/prmpt/mytags.txt, then reference it as __pmp/prmpt/mytags__
Remove a wildcard: delete the __...__ token from the WildcardProcessor text field in the workflow
Disable randomization: replace the __wildcard__ token with a fixed string

Required Custom Node

ComfyUI-TagForge (includes the WildcardProcessor node and the __pmp/prmpt/* wildcard set) — huchukato/ComfyUI-TagForge

The wildcard files ship with the custom node. If a wildcard resolves to empty, the custom node is missing or the wildcard folder is not installed.

🐳 Docker / Cloud Ready

OneClick RunPod Template

Prefer a ready-to-go environment? Use the OneClick - ComfyUI - MiniMax H3 Turbo - Qwen3VL RunPod template:

Docker image: huchukato/comfyui-qwenvl-runpod:cu13-mmh3
Base: huchukato/comfyui-base:cu130
All custom nodes pre-installed
ComfyUI Args: --highvram --disable-auto-launch --fast fp16_accumulation --enable-triton-backend --force-fp16
All 5 Turbo workflows auto-downloaded at boot
Models auto-downloaded at first boot (~96 GB including INT8 diffusion, INT8 Heretic text encoder, 10Eros INT8 and turbo/reference LoRAs; persistent)
ComfyUI tracks the current upstream master branch by default
Comfy Kitchen Attention, FP16 accumulation, high-VRAM mode
TensorRT upscaling + RIFE interpolation (stable defaults: Upscaler 2/2, RIFE 1/1)
SOL-ATTN acceleration active in all modes (tau 1.30 turbo, 1.00 native)

Access: ComfyUI :8188 · JupyterLab :8888 · FileBrowser :8080 (user admin / password adminadmin12) · SSH ssh root@pod-ip

📖 README & instructions

ComfyUI Args (pre-configured)

--highvram
--disable-auto-launch
--fast fp16_accumulation
--enable-triton-backend
--force-fp16

🚀 Why Choose ComfyUI-QwenVL-Mod + MiniMax H3?

🎬 For Content Creators

Native audio: Video and audio in one pass — no separate MMAudio needed
Multilingual: Write in any language, Qwen3.5 handles translation
Professional: Official MiniMax H3 prompt format with camera vocabulary and speaker tags
Quality: 768p native, TensorRT upscale to higher resolution

🔥 For NSFW Content

Explicit: Uncensored generation with dedicated NSFW presets
9 presets: 3 base 🎬 + 3 FL2VA 🔄 + 3 R2VA 🎞️ — each tuned for its mode
Detailed: Rich scene descriptions with explicit diegetic soundscape
Natural: Realistic progression, consistent characters
Audio: Native moans, breaths, skin contact, ambient sounds

⚡ For Power Users

Customizable: Easy to modify presets and system prompts
Extendable: Add your own Qwen3.5 models (GGUF or HF)
Integrable: Works with existing ComfyUI setups
Optimized: Comfy Kitchen Attention, FP16 accumulation, high-VRAM mode, smart caching
Multi-reference: image2 input for FL2VA and R2VA workflows

🌟 What Makes This Special?

First: Complete MiniMax H3 workflow pack with Qwen3.5 auto-prompting
Native audio: No separate audio node — MiniMax H3 does it all
4 Turbo workflows: T2VA, I2VA, FL2VA, R2VA — covers all MiniMax H3 modes
Multi-reference: Qwen3.5 analyzes the first two connected reference images; additional MiniMax references can be described with [P3] / [P4] tags
TensorRT: Built-in upscaling and frame interpolation
9 NSFW presets: Dedicated presets for each mode with correct prompt structure
Multilingual: Any input language, auto-translated and formatted
Ready: Works out-of-the-box with included workflows

🎯 What's New in v2.6.0

⚡ Pure INT8 ConvRot — Default

✅ Default diffusion models: pure INT8 ConvRot (minimax_h3_fl2va_pruned_int8_convrot / minimax_h3_ref2va_pruned_int8_convrot) from Comfy-Org
✅ NVFP4 removed — visible quality loss; speed via turbo LoRA + sol-attn + triton backend
✅ Runs on any modern NVIDIA GPU
✅ Text encoder: uncensored Heretic INT8 ConvRot (~26.4 GB)

🔧 SOL-ATTN Integration

✅ All Turbo workflows include SOL-ATTN nodes (sparse attention + fused modulation + chunked FFN)
✅ Tau per mode: 1.30 turbo, 1.00 native
✅ Spectrum and DiffAid removed from the stack (see 2026/09/16 update)
✅ Turbo LoRA linked from preset to subgraph in all workflows

⚡ Turbo Step Standardization

✅ All Turbo workflows standardized to 8 steps (was 6 for T2VA/R2VA)
✅ Consistent minimax_h3_fl2v_turbo_8step_v1.0_768p_comfyui_bf16 LoRA across all workflows

📦 TensorRT Batch Size

✅ Stable defaults: RIFE loader/runner 1/1, Upscaler loader/runner 2/2
✅ Upscaler 2/2 verified on RTX PRO 6000 Blackwell
⚠️ RIFE batch >1 currently fails during interpolation even when the engine builds; keep it at 1/1
⚠️ Upscaler batch 4 fails to build with TensorRT 10.15 on Blackwell sm_120

🧹 Cleanup

✅ Removed was-node-suite from MiniMax Dockerfile (not used by MiniMax workflows)
✅ Removed ComfyUI-Frame-Interpolation from MiniMax (uses RIFE TensorRT instead)
✅ Removed KJNodes from provisioning (baked into RunPod base image)
✅ ComfyMathExpression is built into ComfyUI core — no custom node needed

🧠 Qwen3.5 Thinking Fix

✅ /no_think prefix for Qwen3.5 models (enable_thinking deprecated in recent llama.cpp)
✅ Broadened architecture detection (qwen35, qwen35moe, qwen35_vl)
✅ Works across both HF and GGUF nodes

📦 Workflow Organization

✅ Moved workflows to minimax/ folder
✅ Renamed FLF to FL2VA (clearer naming)
✅ Added Civitai documentation

📋 Credits

MiniMax H3 — MiniMax · Comfy-Org/MiniMax-H3
ComfyUI — comfyanonymous/ComfyUI
QwenVL-Mod — huchukato/ComfyUI-QwenVL-Mod
Qwen3.5 — Qwen Team / Alibaba
INT4 models — Merserk/MiniMax-H3-INT4-ConvRot
TensorRT RIFE / Upscaler — huchukato
Block Sparse Attention — ComfyUI core
VideoHelperSuite — Kosinkadink
Easy-Use — yolain
PerfectVideoResolution — huchukato

📄 License

Workflows are released under the same license as the underlying models and custom nodes. See each repository for details.

MiniMax H3 model weights: Comfy-Org/MiniMax-H3 — MiniMax H3 Community License.

Built with ❤️ for the ComfyUI community

Files with SHA256: b88c529555bbb3a4f9b5008f5376fec9adf6197f6c20b887adf3c9c8c5138eb8