━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

✨ Krea2 → MiniMax H3 Motion Suite — Text Prompt to Animated Video, One Graph

ComfyUI · Krea-2 Community License + MiniMax H3 Community License · Turbo T2I → H3 I2V (int8) with native audio

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

Type a SUBJECT and a POSE, press Queue, get a still image and an animated video with native audio — no separate tools, no manual image export/import between workflows. This graph fuses Krea2 Turbo (fast DiT text-to-image) with MiniMax H3 (image-to-video with native audio generation) in one funnel: Krea2 renders the starting frame, a VRAM-unload bridge clears it from memory, built-in QwenVL prompt enhancement turns your short SUBJECT + POSE into a rich photographic prompt (and, separately, an automatic motion caption from the rendered frame), then H3 int8_convrot animates it — sound included. This is the sister release to my Krea2-Turbo Dual and Wan2.2 I2V Funnel workflows: same house pattern (cleanGpuUsed bridge, QwenVL auto-prompt), now with H3's native audio track.

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

🔞 Mature Content Disclosure

This workflow's QwenVL prompt-enhancer system prompts are tuned for adult glamour, intimate, and nude photography output (both the SUBJECT prompt expander and the H3 motion-caption generator). All example showcase images/clips depict fictional adult (18+) subjects generated entirely by AI — no real people, no likeness of any real individual. Individual showcase items are marked with their own explicit content rating (X/Mature) per Civitai's per-image rating system rather than flagging the entire listing page as mature.

Users are responsible for following Civitai's Terms of Service and their local jurisdiction's laws regarding generated adult content when using this workflow.

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

✨ Features

✅ One Graph, Zero Handoff — SUBJECT + POSE prompt in, still image AND animated video with audio out. Krea2 generates the frame internally and hands it straight to H3 — no LoadImage node juggling between separate workflows.

✅ SUBJECT + POSE Boxes with Built-In Examples — two clearly labeled, resizable text boxes (SUBJECT, POSE/ACTION) sit at the top of the graph, each with an example-list Note right beside it — copy an example in, or write your own. No hunting through the graph for what's tunable.

✅ VRAM-Safe Double Bridge — easy cleanGpuUsed unloads Krea2 before QwenVL's motion-caption pass loads, then unloads QwenVL again before H3 loads — the three models never co-reside in VRAM.

✅ Native Audio — H3 generates a synchronized stereo audio track alongside the video, decoded via a dedicated audio VAE — no separate audio-gen step.

✅ Dual Prompt Enhancement — QwenVL expands your SUBJECT + POSE into a ~90-word photographic prompt for the still, and separately auto-captions motion/camera/atmosphere from the rendered frame for the video (or write your own MOTION text manually via a switch).

✅ Mode A / Mode B Switch — Mode A: SUBJECT+POSE boxes drive generation (default). Mode B: drop a reference image instead, QwenVL describes it and Krea2 regenerates inspired by it.

✅ Dual Orientation — one switch flips both stages together: 1920×1088 landscape (16:9) or 1088×1920 portrait (9:16) — Krea2 and H3 always stay in matching aspect.

✅ Auto / Manual Motion Switch — default auto-captions motion from the frame; flip one switch to type your own MOTION text instead.

✅ Krea-2 Original Architecture — NOT FLUX-derived; DiT 12.9B model from Krea, FP8 quant by AlperKTS.

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

📦 Required Models (8 files, ≈52 GB total)

(+1 file, ~592 MB, for the Turbo 8-step variant used in all showcase clips — see "v1.1 — Turbo 8-Step Variant" section below for details)

Krea2 T2I stage (~10 GB):

• krea2_turbo_fp8.safetensors (~7 GB) — Krea2 Turbo DiT 12.9B UNet, FP8 quant by AlperKTS

• qwen3vl_4b_fp8_scaled.safetensors (~2–3 GB) — Krea2 text + vision encoder (Qwen3-VL-4B, FP8)

• qwen_image_vae.safetensors (~1 GB) — Krea2 image VAE

MiniMax H3 I2V stage (~42 GB, int8_convrot variant):

• MiniMax_H3_FL2VA_pruned_int8_convrot.safetensors (21 GB) — H3 diffusion UNet, int8_convrot quant

• qwen3vl_32b_minimax_h3_int4_convrot.safetensors (15 GB) — H3 text encoder (Qwen3-VL-32B, int4)

• minimax_h3_video_vae_fp16.safetensors (5.21 GB) — H3 video VAE

• minimax_h3_audio_vae_fp32.safetensors (605 MB) — H3 audio VAE (native audio decode)

Prompt Enhancement (auto-downloaded on first use):

• Qwen3-VL-2B-Instruct (~2.5 GB, auto-cached) — VLM enhancer for SUBJECT/POSE→prompt and frame→motion caption; downloaded automatically by ComfyUI-QwenVL on first run, not manually placed

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

⬇️ Download Links

📁 ComfyUI/models/unet/Krea2/

• krea2_turbo_fp8.safetensors — https://huggingface.co/AlperKTS/Krea2_FP8

📁 ComfyUI/models/clip/ (or text_encoders/)

• qwen3vl_4b_fp8_scaled.safetensors — https://huggingface.co/AlperKTS/Krea2_FP8

📁 ComfyUI/models/vae/

• qwen_image_vae.safetensors — https://huggingface.co/AlperKTS/Krea2_FP8

📁 ComfyUI/models/unet/MiniMaxH3/

• MiniMax_H3_FL2VA_pruned_int8_convrot.safetensors — https://huggingface.co/Abiray/Minimax-H3-nvfp4-INT4-INT8-Convrot

📁 ComfyUI/models/text_encoders/

• qwen3vl_32b_minimax_h3_int4_convrot.safetensors — same repo, text_encoders/ folder

📁 ComfyUI/models/vae/ (H3, separate from Krea2 VAE above)

• minimax_h3_video_vae_fp16.safetensors — same repo, vae/ folder

• minimax_h3_audio_vae_fp32.safetensors — same repo, vae/ folder

⚠️ Filenames above are exactly what the workflow JSON expects in its loader nodes — match them exactly, or update the loader node if your local copy is named differently. The H3 repo also hosts nvfp4 and mixed-int4/int8 variants — this workflow is built and tested specifically against the *int8_convrot** files listed above; other variants will need loader-node edits.*

📁 ComfyUI/models/loras/ (required for the Turbo 8-step variant — see "v1.1 — Turbo 8-Step Variant" section below)

• minimax_h3_turbo_4step_ckpt500_pruned_comfyui.safetensors — https://huggingface.co/drbaph/MiniMax-H3-Turbo-Lora-ComfyUI

Auto-Downloaded Models (no manual placement needed):

• Qwen3-VL-2B-Instruct — automatically downloaded by the ComfyUI-QwenVL custom node on first workflow run (~2.5 GB, cached afterward).

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

🧩 Required Custom Nodes (3 packs)

1. ComfyUI-QwenVL (1038lab / AILab) — provides AILab_QwenVL_PromptEnhancer (SUBJECT+POSE → rich prompt) and AILab_QwenVL (image → motion caption / Mode B description). License: Apache-2.0/BSD.

2. ComfyUI-Easy-Use (vjumpkung fork) — provides easy cleanGpuUsed (VRAM-unload bridge) and easy anythingIndexSwitch (Mode/Orientation/Motion switches).

3. ComfyUI-WAS-Node-Suite (WASasquatch) — provides Text Multiline (the SUBJECT and POSE input boxes). License: MIT.

MiniMax H3's own nodes MiniMaxH3ImageToVideo, MiniMaxH3SigmaShift) are native to ComfyUI ≥0.30.1 — no extra custom node pack needed for those, just an up-to-date ComfyUI install.

Install via ComfyUI Manager (search each pack name, or use "Install Missing Custom Nodes" after loading the JSON).

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

🚀 How to Use

Quick Start:

1. Download all 7 manually-placed model files → place in ComfyUI/models/ (see paths above)

2. Install the 3 custom node packs via ComfyUI Manager, and confirm ComfyUI ≥0.30.1 (for native H3 nodes)

3. Load the workflow JSON into ComfyUI

4. Type your scene into the SUBJECT box (top of graph) — copy one of the example lines from the Note beside it, or write your own

5. Type your pose/action into the POSE / ACTION box beside it — same idea, examples right there

6. Leave Mode Switch = 0 (Mode A), Motion Switch = 0 (auto-caption) for the simplest path

7. Queue → Krea2 renders the still → cleanGpuUsed unloads it → QwenVL auto-captions motion from the frame → cleanGpuUsed unloads QwenVL → H3 animates it with native audio → MP4 lands in your output folder

What's happening under the hood:

- Krea2 T2I: SUBJECT + POSE (combined by StringConcatenate, expanded by QwenVL PromptEnhancer) → EmptySD3LatentImage (1920×1088 or 1088×1920) → KSampler (euler, simple, 10 steps, cfg 1.0) → VAEDecode

- Bridge #1: easy cleanGpuUsed clears the Krea2 stage from VRAM

- Motion caption: QwenVL auto-captions the rendered frame into a ~60–100 word motion/atmosphere description (or your own Manual Motion text, via switch)

- Bridge #2: easy cleanGpuUsed clears QwenVL from VRAM before H3 loads

- Prep: ImageScale to H3's input size (orientation-aware, matches the Krea2/H3 orientation switch)

- H3 I2V: MiniMaxH3ImageToVideo (conditioned on the scaled still + motion caption) → MiniMaxH3SigmaShift (video=12.0, audio=3.0) → SamplerCustomAdvanced (res_multistep, 20 steps, simple scheduler) → VAEDecode (video) + VAEDecodeAudio (audio)

- Output: CreateVideo (24fps, stereo audio) → SaveVideo → H.264 MP4 with synced audio track

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

⚙️ Settings & Parameters

• Krea2 Sampler — euler, simple scheduler, 10 steps, CFG 1.0

• Krea2 Resolution — 1920×1088 (16:9) or 1088×1920 (9:16), via Latent Switch

• H3 Sampler — res_multistep, simple scheduler, 20 steps, SamplerCustomAdvanced

• H3 Sigma Shift — video=12.0, audio=3.0 (defaults, matched to bench-confirmed production settings)

• H3 Resolution / Length — 1280×704 (16:9) or 704×1280 (9:16), 124 frames @ 24fps (~5.17s), via H3 Orientation Switch

• Mode Switch — 0 = Mode A (SUBJECT+POSE boxes) / 1 = Mode B (reference image)

• Motion Switch — 0 = Auto (QwenVL captions the frame) / 1 = Manual (type your own MOTION text)

• Latent Switch / H3 Orientation Switch — 0 = Landscape / 1 = Portrait (keep both switches matched)

• Negative Prompt — curated quality filter, includes anatomy guards (bad anatomy, extra limbs, malformed hands)

• VRAM Bridges — easy cleanGpuUsed ×2 (do not remove — these keep Krea2 / QwenVL / H3 from co-residing in VRAM)

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

✍️ Writing SUBJECT + POSE (Realism Recipe)

Two boxes, two jobs: SUBJECT = who + what they're wearing + the scene; POSE / ACTION = how they're positioned. QwenVL combines and expands both into one photographic prompt automatically.

SUBJECT examples (copy one in, or use as a template):

• petite young Thai woman, sheer lace lingerie

• curvy mature Latina woman, silk robe half-open

• athletic Korean woman, oversized shirt no pants

• tall Scandinavian blonde, sheer slip dress

• curvy Black woman, satin bralette and shorts

• petite Japanese woman, open kimono

• mixed-race woman, wet white t-shirt

• MILF-type woman, unbuttoned denim shirt

POSE / ACTION examples:

• relaxed candid pose on a bed, natural expression

• leaning against window, soft gaze at camera

• sitting cross-legged on floor, playful smile

• stretching on couch, looking over shoulder

• standing in doorway, hand on frame, confident stance

• kneeling on bed, hair falling over shoulder

• walking toward camera, candid mid-step

• lying on side, propped on elbow, dreamy expression

Mix and match freely — SUBJECT and POSE are independent boxes, so any combination works. Short inputs are fine; QwenVL fills in camera/lighting/skin-realism detail automatically.

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

🎨 Showcase

5 clips below, all generated with the v1.1 Turbo variant — Portrait 704×1280, native audio kept, motion auto-captioned by QwenVL (no manual motion text).

⚠️ Disclosure: showcase stills used two extra personal LoRAs stacked on the Krea2 checkpoint (a character LoRA + a skin-realism LoRA) — not included in this workflow. The shared JSON is base-checkpoint only; your results will look great but slightly less hyper-real on skin texture than these samples. Full per-clip prompts + settings in GENERATION_DATA_v1.1.txt.

1. Night city street — elegant black evening dress, rain-slicked street at night, neon signs reflected in puddles

2. Golden-hour beach — flowing white sundress, barefoot walk along the shoreline, warm backlit sun

3. Studio glam — fitted deep red satin top, moody studio lighting, high-fashion editorial glamour

4. Boudoir lace — black lace lingerie, satin-sheet bed, dim warm bedroom lighting

5. Neon slipdress — fitted red silk slip dress, rain-streaked window at night, city neon glow

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

💡 Performance Tips

• Minimum Spec — 16 GB VRAM and ~32 GB system RAM. The double cleanGpuUsed bridge offloads models to system RAM between stages; low-RAM systems may thrash during handoff.

• VRAM — Tested on RTX 5080 16GB — peak ~14.9–15.2 GB across real batch runs (int8_convrot, native resolution, no upscale). Treat 16GB as a hard minimum, not a comfort margin.

• Cold Start — On first workflow run, Qwen3-VL-2B-Instruct (~2.5 GB) auto-downloads and caches. Plan an extra 30–60 seconds the first time only.

• int8_convrot Quant — chosen for broad GPU compatibility (works on 30/40-series, not just Blackwell/50-series); if you're on a Blackwell card, the same HF repo's nvfp4 variant is a smaller/faster drop-in (swap the two H3 loader nodes).

• SUBJECT / POSE Split — keep appearance/scene in SUBJECT and positioning in POSE; blending them into one box tends to confuse the enhancer.

• Motion Auto-Caption — reads the rendered still, so a clearer, less cluttered SUBJECT/POSE composition tends to produce a cleaner motion caption too.

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

📝 Notes & AI Disclosure

• AI-Generated Content — All example outputs are AI-generated by Krea2 Turbo + MiniMax H3. Respect local AI disclosure laws when publishing.

• Hardware — Tested on RTX 5080 16GB — peak VRAM ~14.9–15.2GB across real batch runs (see Performance Tips above).

• Configuration Only — no model weights embedded in the JSON; download all 7 manually-placed files separately from the sources listed above.

• Workflow Reuse — feel free to modify, share, and fork.

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

🔗 Also check out

🔗 Sister workflow: [Krea2-Turbo Dual](https://civarchive.com/models/2736816) — the still-image engine this funnel is built on.

🔗 Sister workflow: Wan2.2 I2V Funnel — a different Krea2→video funnel (Wan2.2-I2V-A14B, RIFE-smoothed, no native audio) if you want a lighter/faster video stage without sound.

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

## ⚡ v1.1 — Turbo 8-Step Variant — 2.3× Speed Gain

Use the Turbo LoRA to cut generation time by more than half. A turbo-distilled variant cuts H3 render time from ~509 seconds to ~217 seconds per 5-second clip on RTX 5080 Laptop 16GB — a 2.3× speedup — via one LoraLoaderModelOnly node stacking a pre-trained H3 distillation LoRA. Steps drop from 20 to 8 with no quality compromise on motion, identity, or audio.

What's included:

- New workflow file: funnel_krea2_h3_v1_turbo.json (identical graph to v1.0, one extra LoRA node)

- Requires downloading the MiniMax H3 Turbo LoRA (Apache-2.0):

- Source: [drbaph/MiniMax-H3-Turbo-Lora-ComfyUI](https://huggingface.co/drbaph/MiniMax-H3-Turbo-Lora-ComfyUI) on HuggingFace

- File: minimax_h3_turbo_4step_ckpt500_pruned_comfyui.safetensors (~592 MB)

- Placement: ComfyUI/models/loras/

- No additional custom nodes needed — uses the same 3 node packs as v1.0

Measured Performance (controlled bench, RTX 5080 Laptop 16GB, same seed, same input):

- 124-frame clip (704×1280 portrait, 5.17 seconds @ 24fps):

- v1.0 (20 steps, no LoRA): 508.8 seconds

- v1.1 Turbo (8 steps + LoRA strength 1.0): 217.0 seconds

- Speedup: 2.3× (57% faster)

- 8 steps verified in our tests; the LoRA author reports 6–8 as the working range; 4–5 steps possible but untested

- Quality verdict (eye-check, same-seed comparison): motion, identity and audio hold up identically. Turbo output shows slightly sharper, more textured skin and stronger camera push-in vs v1.0's softer, more film-like look — a style shift, not a defect. Pick based on your aesthetic preference.

When to use:

- Turbo variant: speed matters, want the sharper skin/dynamic camera look (previews, iterations, long-clip batches)

- v1.0 original: prefer the softer, more cinematic look, unlimited render time (final outputs, hero frames)

Reverting: to switch back to v1.0, simply bypass the LoRA node and set H3 steps back to 20.

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

⭐ Found this useful?

• Like if it saved you time animating stills with sound

• Comment your results — I read every one

• Follow for new ComfyUI workflows, all tested on 16 GB VRAM

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

⚖️ Model Attribution & Licensing

Krea-2 Turbo (AlperKTS FP8 Quant)

• License: Krea-2 Community License — https://www.krea.ai/krea-2-licensing

• Commercial use: OK if total annual revenue < $1,000,000 USD

MiniMax H3 (MiniMax, quant by Abiray)

• License: MiniMax H3 Community License — https://huggingface.co/Abiray/Minimax-H3-nvfp4-INT4-INT8-Convrot

• "Excluded Territories" clause governs where the model is run (EU, UK, Republic of Korea, USA) — running this workflow in an unrestricted territory and selling/distributing outputs is not restricted by this clause; MiniMax claims no rights over generated outputs.

Supporting Components — all Apache-2.0

• Qwen3-VL-4B (Krea2 text/vision encoder) • Qwen Image VAE • Qwen3-VL-32B int4 (H3 text encoder) • Qwen3-VL-2B-Instruct (prompt enhancer model)

NSFW generation permitted under both licenses.

ComfyUI Custom Nodes

• ComfyUI-QwenVL (1038lab) — Apache-2.0/BSD

• ComfyUI-Easy-Use (vjumpkung fork) — per upstream repository

• ComfyUI-WAS-Node-Suite (WASasquatch) — MIT

Workflow JSON — original work, free to use, modify, and redistribute.

Full attribution detail (sources, license text, verification dates) is maintained separately in this listing's ATTRIBUTION.md. All example outputs are AI-generated. Model weights remain the property of their respective owners; download separately from the official sources above.