MiniMax H3 TFL2VA — Workflow Guide

92 nodes · 10 functional groups · 122 links · native video and stereo-audio generation

Text-to-video-and-audio, first-frame image-to-video-and-audio, and first/last-frame interpolation
Optional local AI prompt expansion
Modular SOL-Attention, SPECTRUM, and DIFF-AID model patching
Optional 2× TensorRT frame upscaling
H.264 MP4 output with embedded workflow and generation metadata

What Is This?

This is a focused MiniMax H3 workflow for generating video and native stereo audio together in ComfyUI. Depending on which image controls are enabled, the same graph supports three useful modes:

T2VA — text to video and audio, with no reference image.
I2VA — a prompt plus a starting image to video and audio.
TFL2VA — a prompt plus starting and ending images to video and audio.

The workflow keeps configuration, images, prompts, models, sampling, upscaling, and saving in separate groups. Eclipse Set/Get nodes carry values between them like wireless connections, so the canvas remains readable and optional stages can fall back to the last available result.

MiniMax H3 models the visual and audio streams jointly. Describe dialogue, sound effects, ambience, and music in the same prompt as the action and camera direction; the soundtrack is not added as a separate post-process.

The Complete Flow

Folder / Video Settings
        ↓
Optional starting and ending images
        ↓
Manual prompt or Smart LM expanded prompt
        ↓
MiniMax H3 model + optional LoRAs
        ↓
SOL-Attention → DIFF-AID → SPECTRUM
        ↓
MiniMax H3 conditioning and AV latent sampling
        ↓
Video VAE decode + Audio VAE decode
        ↓
Optional TensorRT 2× frame upscale
        ↓
H.264 MP4 with native audio and embedded metadata

The saved workflow sends the patched model through SOL-Attention, then DIFF-AID, then SPECTRUM. This lets SPECTRUM see DIFF-AID's compatibility metadata and protect its forecast history correctly.

Before You Run It

Update ComfyUI

This workflow uses MiniMax H3 nodes from ComfyUI core, including MiniMaxH3ImageToVideo and VAEDecodeAudio. Update ComfyUI if those nodes are missing.

Desktop and cloud releases may lag behind the newest core implementation. A fully updated manual installation may therefore support H3 before a stable desktop build does.

My Tested 16 GB Environment

This is the configuration I use and have tested:

ComfyUI: 0.34.0
Frontend: 1.49.6 with the two compatibility fixes described in ComfyUI Frontend 1.46.2 / 1.48.7 / 1.49.6 — Linux & Windows
NVIDIA PyTorch build for a 16 GB GPU: CUDA 13.0 (cu130)

For a 16 GB NVIDIA GPU or lower, use a PyTorch cu130 build so ComfyUI can use comfy_kitchen's optimized H3 CUDA operations. Without that optimized path, this workflow may run out of VRAM.

The ComfyUI Installation Script for Linux can install a matching PyTorch environment, check the CUDA toolchain, install the matching versioned CUDA compiler and development packages when needed, and build SageAttention 2 on supported Linux/NVIDIA systems.

This is my known-working configuration, not a guarantee that every 16 GB card (or lower) can run every resolution and duration. Begin with a 5-second render and keep TensorRT Upscale bypassed until the base workflow succeeds.

Required External Repositories

These workflow stages come from separate custom-node repositories:

ComfyUI Eclipse — workflow routing, folder controls, image preparation, mode toggles, metadata, and video saving.
ComfyUI Smart Model Loader — H3 model, text encoder, video/audio VAE, LoRA, sampling, and memory configuration.
ComfyUI SmartLLM — multimodal prompt generation through its supported local or remote backends.
ComfyUI-sol-attn — the active scheduled SOL-Attention, fused-modulation, and chunked feed-forward group.
KJNodes — sigma visualization and the optional SageAttention fallback chain.
ComfyUI Upscaler TensorRT Auto — optional TensorRT frame upscaling.
ComfyUI-DiffAid-Patches — the MiniMax H3 DIFF-AID patch group.
ComfyUI-Spectrum-MiniMax-H3 — the MiniMax H3 SPECTRUM patch group.

If SPECTRUM does not appear in ComfyUI Manager search, install it directly from its GitHub repository:

cd ComfyUI/custom_nodes
git clone https://github.com/xmarre/ComfyUI-Spectrum-MiniMax-H3

Restart ComfyUI after installing or updating custom nodes. All listed repositories provide nodes used by the complete graph and should be installed to avoid missing-node warnings. SmartLLM is optional at runtime when Auto Prompt is disabled; the DIFF-AID and SPECTRUM groups can be bypassed for a native baseline; and Upscale can be bypassed to save the original decoded frames.

Required Models

The H3 model download bundle and the Smart Model Loader template are attached to the ComfyUI Smart Model Loader changelog article. Use that bundle for the prepared setup, or download the individual files below and place them in their matching ComfyUI model folders.

Diffusion model — ComfyUI/models/diffusion_models/

minimax_h3_fl2va_pruned_int8_convrot.safetensors

Text encoder — ComfyUI/models/text_encoders/

qwen3vl_32b_minimax_h3_nvfp4_awq.safetensors

VAEs — ComfyUI/models/vae/

minimax_h3_video_vae_fp16.safetensors
minimax_h3_audio_vae_fp32.safetensors

Optional LoRA Slots

Smart Model Loader provides up to four built-in LoRA slots. Every slot is optional and can be turned on or off independently. The saved workflow includes minimax_h3_turbo_4step_ema_ckpt850_pruned_comfyui.safetensors as a selection in one slot, but all four switches are off.

The separate LoRA Stack provides up to ten additional slots when four are not enough, or when you want to combine several LoRAs in one reusable stack. Its entries are also optional and independently switchable. All ten switches are off in the publication workflow, so missing optional LoRA files will not prevent the model from loading.

The first separate-stack slot is prepared for MiniMax H3 Astro NSFW, using the file minimax_h3_Astro nsfw.safetensors at model strength 0.70. Install that LoRA before enabling the slot, or replace it with another LoRA you own.

Select a Reference Image

The publication workflow resets the starting-image loader to input/example.png instead of a local output path. Select your own image before queuing when you want first-frame guidance. The ending-image branch is bypassed by default.

Group-by-Group Guide

Folder / Video Settings

This is the workflow's central output, resolution, frame-rate, and duration control.

The saved defaults are:

Mode: Video Mode
Resolution: 736 × 1280 portrait
Divisibility: 32 pixels
Duration: 10 seconds
Frame rate: 24 fps
Output root: video
Date folder: enabled with %Y-%m-%d

The Math Expression node converts seconds and frame rate into the frame count H3 requires. It rounds the requested duration and then snaps upward to H3's valid 17k + 5 frame grid.

Examples at 24 fps:

5 seconds becomes 124 frames, approximately 5.17 seconds.
10 seconds becomes 243 frames, approximately 10.13 seconds.
15 seconds becomes 362 frames, approximately 15.08 seconds.

The native node describes approximately 124–362 frames as the trained range. Longer clips are accepted by the node but are untested and require much more memory.

For a first test, keep the default 736 × 1280 resolution and use 5 seconds. Increase duration only after the full model, patch, decode, and save path succeeds.

The visible width, height, and length widgets on MiniMaxH3ImageToVideo are connected inputs. The values from Smart Folder and Math Expression override those saved widget values during execution.

Load Images

Two parallel image branches provide H3's optional keyframes:

Starting Image supplies frame 0 and is enabled by default.
Ending Image supplies the final frame and is bypassed by default.

Each branch loads an image, resizes it to the active video canvas, and publishes it as ref_image_1 or ref_image_2. The resize uses Lanczos, center crop, and dimensions divisible by 32.

Use the Fast Mode Toggle to select the generation mode:

Text only: turn both Starting Image and Ending Image off.
First-frame animation: turn Starting Image on and Ending Image off.
First-to-last interpolation: turn both on and select a valid image in each loader.

For first/last-frame generation, H3 fixes the first reference at frame 0 and the second at the final valid frame. The model generates the motion, intermediate composition, and synchronized audio between them.

A last frame is a strong constraint, not merely a style reference. If motion becomes unnatural, simplify the prompt, choose endpoints with compatible framing, or return to first-frame-only mode.

Both active references are also joined and passed to the optional Smart LM prompt stage, allowing the language model to describe a transition informed by both endpoints.

Prompt

The Positive Prompt node is the main user input. A strong H3 prompt normally combines:

Subject and action
Shot sequence and timing
Camera motion
Lighting and visual style
Spoken dialogue
Sound effects and ambience
Music or emotional tone

Example structure:

integrated_multimodal_description: Live-action cinematic realism with fine
35mm grain, warm evening light, and deep blue shadows. [Shot 1] On a quiet
rooftop, Mara (S1) walks toward the camera as it slowly pulls back. Wind moves
her coat and the city glows behind her. Mara (S1) smiles and says:
<d>[English] We made it just before sunset.</d> [Shot 2] At 00:03.000, cut to a
close side profile as she turns toward the skyline. The camera makes a slow,
small push-in while she exhales and watches the last sunlight fade.

overall_soundscape: Soft rooftop wind, distant traffic, fabric rustling, quiet
footsteps on concrete, and Mara's natural close-miked voice.

non_diegetic_music: A restrained cinematic ambient score with warm analog pads
and a sparse piano motif.

integrated_multimodal_description defines the visual style, subjects, actions, camera work, shot sequence, and synchronized dialogue. Use [Shot 1], [Shot 2], and timestamps to make changes of shot explicit. Keep a speaker identifier such as (S1) consistent, and wrap spoken lines in <d>[Language] ...</d> tags. overall_soundscape describes sounds that belong to the scene, while non_diegetic_music describes the separate score.

Auto Prompt

Auto Prompt is enabled by default. Write a short description of what should happen; the MiniMax H3 Scene 5s SmartLLM task uses it together with any active reference images to generate a detailed prompt in the audiovisual format MiniMax H3 expects. The generated result has first priority in the prompt switch. If Auto Prompt is disabled, the original manual prompt passes through unchanged, so write the complete visual and audio prompt yourself.

SmartLLM includes multiple tasks created specifically for MiniMax H3. Select the H3 task that best matches the type of scene or prompt you want to generate; MiniMax H3 Scene 5s is the task selected in this workflow, not the only H3 option.

The saved task's 5s label does not set the rendered duration. Smart Folder remains the source of truth for duration.

The saved configuration selects huihui_ai-qwen3.5-abliterated-9b-Claude-Ollama, but Auto Prompt is not limited to Ollama. You can select another backend and model supported by SmartLLM, including Transformers and compatible Docker-hosted services.

Prompt Review

The selected prompt passes through Prompt Review, a Show Any Stop node. Its Stop control is off by default, so it displays the prompt and continues. Turn Stop on when you want the queue to halt after prompt construction for inspection.

When comparing patch settings, keep Auto Prompt off or keep its seed and output fixed. Otherwise prompt rewriting can change between tests and invalidate the comparison.

Model Loader

Smart Model Loader loads the H3 diffusion model, Qwen text encoder, video VAE, and audio VAE as one reusable configuration. It also publishes the model metadata used later by the MP4 generation-data block.

Important saved settings include:

Diffusion model: minimax_h3_fl2va_pruned_int8_convrot.safetensors
H3 sampling method: MiniMax H3
Video shift: 12
Audio shift: 5.5
Sampler: Euler
Scheduler: Simple
Steps: 20
CFG: 1.0
Denoise: 1.0

LoRA Application

The separate Eclipse LoRA Stack extends the four built-in Smart Model Loader slots with up to ten more independently switchable LoRAs. You can use either set by itself or combine LoRAs across both. The separate stack is applied to the diffusion model before any optimization or conditioning patch. It runs in model-only mode, so it does not patch the text encoder.

The workflow also combines the enabled LoRA names from Smart Model Loader and the separate LoRA Stack before passing them to Generation Data. This records both sets in the saved video's embedded metadata instead of reporting only the loader's first four slots.

The loader's internal Turbo LoRA is disabled. If you enable a 4-step Turbo LoRA, reduce steps only after confirming the settings recommended for that LoRA. Combining few-step LoRAs with SPECTRUM changes the quality/speed tradeoff more aggressively than SPECTRUM alone.

Optional KJ Sage Fallback Chain

The Model Loader group contains two connected KJNodes patches:

Patch Sage Attention KJ — a global SageAttention override.
MiniMax H3 Memory Efficient Sage Attention Patch — replaces H3 self-attention directly to reduce peak VRAM.

The updated workflow wires them in the recommended position after all loader and separate-stack LoRAs, but both nodes are bypassed by default. The active route is:

Smart Model Loader with its enabled LoRAs
→ separate LoRA Stack Apply
→ Patch Sage Attention KJ
→ MiniMax H3 Memory Efficient Sage Attention Patch
→ Set_MODEL
→ SOL-Attention

Because both Sage nodes are bypassed, the model currently passes through them unchanged. To use the complete Sage fallback chain, switch both nodes from Bypass to Always without changing their order.

Keeping all LoRAs before the attention patches matches the documented SOL-Attention stack:

LoRA Stack Apply
→ Patch Sage Attention KJ
→ MiniMax H3 Memory Efficient Sage Attention Patch
→ Set_MODEL

The global Sage patch provides the general attention backend. The H3 memory-efficient patch then replaces MiniMax H3 self-attention specifically. With SOL downstream, the H3 Sage path becomes SOL's fallback when SOL does not activate for a shape, dtype, token count, or kernel condition. Putting the H3 Sage patch after SOL would instead shadow SOL and prevent it from running.

KJNodes labels its H3 memory-efficient Sage patch experimental and requires a current SageAttention installation. Treat this as an optional VRAM fallback, not a mandatory quality enhancement.

Model Patcher Groups

The three patcher groups are independent. Eclipse priority routing selects the most downstream active model without requiring visible cross-group noodles:

Base MODEL
→ MODEL_SOL
→ MODEL_DIFF
→ MODEL_SPECT
→ Sampler

If a group is bypassed, the next group falls back through the earlier published models. DIFF-AID checks MODEL_SOL and then the base MODEL; SPECTRUM checks MODEL_DIFF, then MODEL_SOL, then the base MODEL; and the sampler checks MODEL_SPECT, MODEL_DIFF, MODEL_SOL, then the unpatched base MODEL.

The physical left-to-right position of these groups does not control execution. Their Get/Set variable priority and model links do.

These groups do different jobs:

SOL-Attention changes how attention and feed-forward work are executed to improve speed or memory use.
SPECTRUM predicts selected denoiser features so some full H3 transformer evaluations can be skipped.
DIFF-AID modifies language-conditioning strength in selected transformer blocks to experiment with prompt adherence.

Only SPECTRUM is primarily a step-skipping accelerator. DIFF-AID is not a speed patch.

SOL-ATTN

The SOL-ATTN group applies three MiniMax-specific optimizations in sequence.

Scheduled SOL Attention

This replaces eligible H3 self-attention calls with a sparse SOL-Attention kernel. Sparse routing is strongest early in denoising, when the structure is loose, and becomes more conservative late in denoising, when fine detail forms.

The saved profile uses:

tau_start = 1.3
tau_end = 0.8
curve = linear
min_tokens = 4096
dense_percent = 0
sink_conditioning = exact_kv
int8_qk = false
int8_pv = false
strict = false

Higher tau routes more key/value blocks through the approximate sparse path, generally favoring speed over fidelity. The falling 1.3 → 0.8 schedule is designed to spend more exact work near the end.

exact_kv keeps H3's packed text, conditioning, reference, and audio key/value blocks exact. This is an important safeguard for prompt adherence and audio synchronization.

Below 4096 packed tokens, on an unsupported shape, or after a non-strict kernel failure, the node falls back to the attention path it received. Both connected KJ Sage nodes are bypassed in the saved workflow, so the default fallback is the base loader's selected attention backend. If the H3 memory-efficient Sage patch is activated before SOL, its H3 attention implementation becomes SOL's fallback instead.

Fused Modulation

MiniMax H3 repeatedly applies adaptive layer-normalization scale/shift and gated residual operations across its 50 DiT blocks. Fused Modulation combines that elementwise work into fewer Triton launches.

This is intended as a numerically exact execution optimization rather than an artistic effect. It does not choose an attention backend and can coexist with SOL or Sage attention.

Do not combine this node with KJNodes' separate “MiniMax H3 Low VRAM Attention” whole-block patch. Both replace the complete H3 block forward and are documented as incompatible.

Chunk FeedForward

H3's feed-forward projection can create a very large temporary activation, especially with the INT8 ConvRot checkpoint used here. Chunk FeedForward splits work along the token dimension.

The saved settings use 2 chunks when the packed sequence has at least 8192 tokens. The mathematics are token-local, so the patch is intended to reduce peak VRAM without changing output. Below the threshold it does nothing.

Use more chunks only when peak MLP memory is the blocker; extra chunking adds launch overhead. This patch is independent of SOL-Attention and remains useful with another attention backend.

When to Use SOL-ATTN

Keep the group active when its supported Triton/CUDA path works on your GPU and you want lower attention cost plus lower feed-forward peak memory. Bypass it when validating the native baseline, investigating kernel compatibility, or comparing output fidelity.

DIFF-AID

The MiniMax H3 DIFF-AID node is an experimental inference-time conditioning patch. It identifies genuine language rows from H3's native segment metadata and applies this controlled enhancement only inside selected transformer blocks:

text rows = text rows + text rows × strength

The saved profile uses:

Blocks 1,13,25,37,50
Strength 0.20
Full sigma window from 0.0 to 1.0
No sigma ramp
No token-position weighting
Conditional rows only

The goal is to test whether modest, sparse text-conditioning reinforcement improves prompt adherence without globally raising CFG. Visual, audio, reference, and target-video rows are left untouched.

This H3 port is not the trained Diff-Aid method from the research paper, and the chosen H3 blocks are an evenly distributed experimental starting set—not a proven quality preset. Stronger values may overconstrain motion, exaggerate literal prompt details, or shift composition.

DIFF-AID and SPECTRUM Order

The saved workflow now uses DIFF-AID's documented Spectrum-compatible route:

SOL → DIFF-AID → SPECTRUM → sampler

Placing DIFF-AID before SPECTRUM lets SPECTRUM consume DIFF-AID's compatibility descriptor, include the patch settings in its cache identity, and force an actual H3 evaluation at a hard activation-window transition when needed.

The saved full-range profile has no interior on/off transition, but correct ordering still matters for cache identity and for any future partial-window settings.

SPECTRUM

SPECTRUM is a training-free, approximate sampling accelerator. On an actual step it runs native H3 and stores the packed target feature state. On a forecast step it predicts that state from previous exact anchors and skips the expensive H3 transformer blocks while the native output and sampler path continue.

For a normal 20-step, one-call sampler, the default schedule commonly produces approximately 11 actual H3 transformer evaluations and 9 forecasts, although fallbacks and safety boundaries can change the count.

The saved profile uses the general defaults:

Enabled with blend weight 0.50
Degree 1 with ridge regularization 0.10
Window size 2.0 and flexible window 0.75
One warm-up step and one exact tail step
History stored in system RAM
Bootstrap first forecast enabled
Offline smoothing replay enabled
Audio blend weight 0.0
Model-aware mode off
SA-PECE policy balanced

audio_blend_weight = 0.0 and offline smoothing replay are conservative audio choices. The capture pass preserves the causal information needed for the soundtrack, followed by a transformer-free replay path intended to avoid the speech and stutter regressions found with direct audio feature mixing.

SPECTRUM is intentionally approximate. The same prompt, seed, and settings can produce a different result than native H3. Use a fixed seed and compare SPECTRUM on versus off for quality-critical work.

SPECTRUM does not replace the attention backend. On actual steps it uses whichever backend arrives from the earlier model chain—SOL in the default active route, or the base attention backend when SOL is bypassed.

Sampler

The sampler group collects the highest-priority active model plus all configuration published elsewhere.

MiniMaxH3ImageToVideo creates two things:

Positive conditioning containing the prompt and optional first/last keyframes
A joint nested latent containing video and stereo-audio streams

SamplerCustomAdvanced denoises both streams together. The result then splits only for decoding:

Video latent → video VAE → image frame batch
Audio latent → audio VAE → stereo waveform

The default runtime sampler configuration is Euler, Simple, 20 steps, CFG 1.0, denoise 1.0. The seed comes from Smart Model Loader.

The sigma preview is diagnostic only. It visualizes the active sigma schedule and does not modify sampling.

Upscale

The decoded frame batch is sent to Auto TensorRT Upscaler when this group is active.

The saved settings are:

Model: 2xLexicaRRDBNet_Sharp
Precision: FP16
Engine batch size: 2
Output resize: 2×

Upscaling changes only the video frames. The audio decoded from the original joint latent is passed unchanged to Save Video.

The upscaler builds a TensorRT engine and requires a compatible NVIDIA/CUDA/TensorRT environment. If the saved model is unavailable, choose another installed supported ESRGAN-style model or bypass the group.

The save stage prefers image_upscale over the base image_output. When Upscale is bypassed, it automatically saves the original decoded frames instead; no rewiring is required.

For the first successful H3 test, bypass Upscale. Enable it only after base video and audio save correctly, because TensorRT engine building is a separate source of installation and VRAM issues.

Video Save

The final stage gathers the chosen frames, decoded stereo audio, frame rate, path, prompt, seed, dimensions, model name, VAE name, sampler, scheduler, steps, and CFG. It merges the enabled LoRA names from Smart Model Loader and the separate LoRA Stack so both sources are represented in the metadata.

The saved output profile is:

MP4 container
H.264 codec
CRF 19
Veryfast preset
24 fps from Smart Folder
Workflow embedded
Generation metadata embedded
Filename prefix MiniMax_H3
Date and time appended to the filename

With the default folder controls, outputs are organized under a date-based folder inside ComfyUI/output/video.

The Trim feature is available, but trim_mode is none in the saved workflow, so it does not shorten the default result.

Quick Start

First-Frame Video With Native Audio

Update ComfyUI and install all required models and custom nodes.
Select your own Starting Image.
Leave Starting Image on and Ending Image off.
Leave all LoRA switches off for the first test.
With Auto Prompt enabled, enter a short description of the scene and select any supported SmartLLM backend and model.
If Auto Prompt is off, write the complete audiovisual prompt manually, including action, camera, dialogue, ambience, and music.
Set duration to 5 seconds for the first test.
Bypass Upscale for the first test.
Queue the workflow.
After a successful base render, enable any installed LoRAs or Upscale and increase duration as desired.

Text-to-Video and Audio

Turn Starting Image off.
Keep Ending Image off.
Enter a complete audiovisual prompt.
Make sure no missing local image remains required by an active image loader branch.
Queue the workflow.

First-to-Last-Frame Video and Audio

Turn Starting Image and Ending Image on.
Select a valid image in both loaders.
Use endpoints with compatible aspect, subject scale, and composition.
Describe the intended transition and the ending action explicitly.
If Auto Prompt is active, review the expanded prompt because it receives both images.
Start at 5 seconds, then increase the duration if the transition needs more time.

Patcher Testing Guide

Model patches can alter performance, memory, or output. Use a fixed seed, prompt, images, duration, resolution, sampler, and LoRA stack for every comparison.

Run these passes in order:

Native baseline: SOL off, SPECTRUM off, DIFF-AID off.
Execution optimization: SOL on, SPECTRUM off, DIFF-AID off.
Forecast acceleration: SOL on, SPECTRUM on, DIFF-AID off.
Conditioning test: SOL on, SPECTRUM off, DIFF-AID on.
Combined stack: SOL on, DIFF-AID before SPECTRUM.

Compare:

Peak VRAM
Total generation time, excluding first-run compilation where appropriate
Prompt adherence
Motion continuity
First/last-frame fidelity
Face and hand stability
Dialogue intelligibility
Sound synchronization
Music and ambience continuity

Do not compare only the final frame. A speed patch can preserve still-image quality while changing motion or audio continuity.

Troubleshooting

Missing MiniMax H3 Nodes

Update ComfyUI. MiniMaxH3ImageToVideo, the H3 model loader support, and VAEDecodeAudio must exist in core.

Missing LoRA Error

All LoRA slots are disabled in the publication workflow. If a missing-file error appears after you enable one, turn that slot off or select a file installed in your own models/loras folder.

Prompt Stage Fails

Auto Prompt requires SmartLLM and a working model on the backend selected in Smart LM Loader. Check that the chosen Ollama, Transformers, or compatible Docker backend is available and that its model is installed or reachable. Turn Auto Prompt off temporarily only when you want to test the manual positive prompt path directly.

Settings Appear to Be Ignored

Many visible sampler widgets have connected inputs. Smart Folder and Smart Model Loader publish the runtime values through Eclipse Set/Get channels. Change the source controls, not the downstream connected widgets.

Out of Memory

Try these changes in order:

Bypass the TensorRT Upscale group.
Reduce duration to 5 seconds.
Reduce resolution while preserving a multiple of 32.
Keep SOL Chunk FeedForward active.
Increase model block swapping or reduce blocks kept on GPU in Smart Model Loader.
Activate the already-wired KJ Sage fallback chain, keeping the two nodes in their saved order.

SOL-Attention Does Not Activate

The scheduled SOL kernel has GPU, Triton, dtype, shape, and token-count requirements. With strict = false, unsupported cases fall back instead of crashing. Check the console log to see whether SOL ran or used its fallback.

SPECTRUM Output Differs at the Same Seed

This is expected: SPECTRUM forecasts selected feature states and changes the denoising trajectory. Bypass it for the native reference result.

Video Saves Without Sound

Confirm that the audio VAE is installed and selected, VAEDecodeAudio completes, the AUDIO Set/Get channel remains active, and Save Video receives its audio input. Include audible events in the prompt; silence or vague audio can also be a model result rather than a wiring fault.

TensorRT Upscale Fails

Test the base video with Upscale bypassed. Then confirm NVIDIA CUDA and TensorRT compatibility, select a supported upscaler model, and allow the engine to build for the active GPU and precision.

LTXVideo Import Errors

LTXVideo is not part of the MiniMax H3 generation path, but an import failure from an installed LTXVideo package can still interfere with ComfyUI startup. If the upstream Lightricks custom node produces import errors in your environment, try my ComfyUI-LTXVideo fork, which includes selected community compatibility fixes.

Keep only one active LTXVideo repository to avoid duplicate node registrations, then restart ComfyUI.

Custom Node Packages Used

Workflow infrastructure

ComfyUI Eclipse — Smart Folder, image loading and resizing, wireless Set/Get routing, mode toggles, prompt review, joining, generation metadata, video saving, and filename construction.
ComfyUI Smart Model Loader — MiniMax H3 diffusion model, text encoder, VAE, audio VAE, sampler settings, memory controls, templates, and LoRA stack.
ComfyUI SmartLLM — optional local multimodal prompt expansion.

Model optimization and output

ComfyUI-sol-attn — scheduled H3 SOL-Attention, fused modulation, and chunked feed-forward patches.
ComfyUI-Spectrum-MiniMax-H3 — approximate H3 feature forecasting and transformer-step acceleration.
ComfyUI-DiffAid-Patches — experimental MiniMax H3 text-conditioning modulation.
KJNodes — the optional SageAttention fallback chain and sigma visualization.
ComfyUI Upscaler TensorRT Auto — optional TensorRT frame upscaling.

Core ComfyUI supplies MiniMax H3 conditioning, joint AV latent sampling, video/audio decoding, the sampler, scheduler, noise, math expression, and basic previews.

Start simple: 5 seconds, one starting image, manual prompt, no upscaler, and all approximate patchers bypassed. Once the native result works, enable SOL, DIFF-AID, SPECTRUM, and TensorRT one stage at a time so every speed, memory, and quality change has a clear cause.

Files with SHA256: 4f5cab0f3f2d61c297eec9a2d1b621fb73aeb2d6993f6f4b9eaf5460ecc0f31c