MiniMax H3 Image-to-Video AUTO — Batch Sequential, Native Audio, QwenVL Auto-Caption (8-Step Turbo)

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
✨ **VID H3 Image-to-Video AUTO — Batch Sequential Workflow with Native Audio & QwenVL Motion Captioning**
ComfyUI · MiniMax H3 Community License · Dual-LoRA Turbo Stack (6-step, ~3× faster sampling)
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

Drop a folder of images → press Queue once per image → workflow auto-iterates, auto-captions motion from each frame with QwenVL, animates with a dual turbo-LoRA stack (official Comfy-Org 8-step distill + lightx2v secondary, 6 steps total), and spits out MP4 video WITH native stereo audio. This is the image-to-video engine: no text prompt, no Krea2 generation stage — just still frames in, up to 10.1s (243-frame) animated clips out. Optimized for 16 GB VRAM with int8_convrot quantization.

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

📋 **Changelog**

**v3 — Dual-LoRA 6-Step** (current)
Speed: dual-LoRA turbo stack (official Comfy-Org 8-step distill + secondary lightx2v LoRA) replaces the single-LoRA v2 setup, running clean at 6 steps (down from 8) — ~3× faster than the original 20-step baseline. Same-seed A/B against v2 showed no quality loss on image or audio. Length: clips now default to 243 frames (~10.1s @ 24fps), up from 124f (~5.2s) in v2. Audio QC hardened this round with a CLAP-based audio/caption similarity pre-filter alongside human-listen review — found and root-caused a real scene-content bias in H3's native audio (beach/swimwear-style scenes can trigger music instead of ambient sound); example clips chosen to avoid it. Seed default switched from fixed to randomize — public giveaway workflow, users now get varied output per run instead of everyone landing on the same result.

**v2.1 — Audio Fix**
Fixed: upscale-path output (⬆️ UPSCALE OUTPUT group) was saving video without audio — the audio VAE decode was never wired into that branch's Create Video node, only the normal-path one was connected. Native audio now comes through on both output modes. Also documented a WAS Node Suite setup gotcha: the caption Save Text File node only writes to paths listed in its whitelist-dirs.list (see Troubleshooting below). Thanks to Civitai commenter @blhll for catching both.

**v2.0 — Auto + Fast VRAM**
Auto-orientation (portrait/landscape now detected from the source image, no manual knob), Turbo LoRA upgraded to v4_step600_ema, lighter VRAM cycle between clips (soft flush replaces full model reload — better on long batches), one-click Normal/Upscale output switch, default clip length raised to 243 frames (~10s), and the whole graph reorganized into 7 numbered guided groups with a first-time setup note built in. New required node: IAMCCS-nodes.

**v1.0 — Initial Release**
First public release — H3 image-to-video with native audio, QwenVL auto-captioning, dual output paths (normal + RTX upscale).

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

✨ **Features**

✅ **Batch AUTO Sequential** — ImageIterator (sequential mode, reset=false) loads one image per Queue press. Press Queue N times for N images, or use Queue (Instant) ×count. No folder juggling, no manual image export/import.
✅ **Auto Motion Caption** — QwenVL reads each input frame and auto-writes a 60–100 word motion description (or manual override via PrimitiveString + index switch 0=auto, 1=manual).
✅ **Dual-LoRA Turbo Stack** — official Comfy-Org 8-step distill LoRA (str 1.0) + secondary lightx2v turbo LoRA (str 0.75) stacked, BasicScheduler set to 6 steps — cuts sampling from 20 steps to 6 (~3× faster) with quality holding up in A/B testing (image+audio both checked).
✅ **Native Stereo Audio** — H3 generates audio alongside video; decode via dedicated audio VAE. To mute, unplug audio input on Create Video node.
✅ **Single Orientation INT** — one knob (0=Landscape 1280×704, 1=Portrait 704×1280) drives H3 latent size and output resolution; no node rewiring.
✅ **Frame Grid 17k+5 Pattern** — 124/141/158/175/192/209/226/243 frames valid; ships defaulted to 243 (~10.1s @ 24fps). ⚠️ Frames ≥175 need ~48 GB system RAM (non-VRAM).
✅ **VRAM-Safe Design** — tested on RTX 5080 16 GB; ComfyUI weight offloading keeps the 21 GB UNet workable on 16 GB cards. int8_convrot quant chosen for broad GPU compat (30/40-series, Blackwell).
✅ **Optional RTX Upscale Side-Branch** — DaSiWa RTX VSR Ultra 1920×1080 (bypassed by default; if red missing-node, just delete — main path runs fine).

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

📦 **Required Models** (5 files, ~42 GB, auto-downloaded ≤2.5 GB)

**MiniMax H3 I2V core (~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)

**Turbo LoRA stack (~1.1 GB, 2 files, both applied):**
• minimax_h3_fl2v_turbo_8step_v1.0_comfyui_bf16.safetensors (str 1.0) — official Comfy-Org 8-step distill LoRA
• minimax_h3_fl2v_lightx2v_turbo_4step_v0.1_comfy.safetensors (str 0.75) — secondary lightx2v turbo LoRA, paired with the above to enable a 6-step BasicScheduler

**Auto-Downloaded (no manual placement):**
• Qwen3-VL-2B-Instruct (~2.5 GB, cached after first run) — VLM motion captioner, auto-fetched by ComfyUI-QwenVL

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

⬇️ **Download Links**

📁 **ComfyUI/models/diffusion_models/**
• 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/**
• minimax_h3_video_vae_fp16.safetensors — same repo, vae/ folder
• minimax_h3_audio_vae_fp32.safetensors — same repo, vae/ folder

📁 **ComfyUI/models/loras/**
• minimax_h3_fl2v_turbo_8step_v1.0_comfyui_bf16.safetensors — https://huggingface.co/Comfy-Org (official H3 turbo distill LoRA)
• minimax_h3_fl2v_lightx2v_turbo_4step_v0.1_comfy.safetensors — lightx2v turbo LoRA, ComfyUI-format release

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

🧩 **Required Custom Nodes** (5 packs)

1. **ComfyUI-QwenVL** (AILab / 1038lab) — motion captioning via AILab_QwenVL node. Apache-2.0/BSD. https://github.com/1038lab/ComfyUI-QwenVL
2. **ComfyUI_Image_Anything** (Kelin) — ImageIterator node for batch sequential loading. ⚠️ **Note:** may not resolve in ComfyUI Manager; install via git clone: `git clone https://github.com/ComfyUI-Kelin/ComfyUI_Image_Anything` into custom_nodes/
3. **ComfyUI-Easy-Use** (yolain) — cleanGpuUsed (VRAM bridge) + anythingIndexSwitch (motion/orientation toggles). https://github.com/vjumpkung/ComfyUI-Easy-Use
4. **rgthree-comfy** — Fast Groups Bypasser (group on/off toggle used in this graph). https://github.com/rgthree/rgthree-comfy
5. **IAMCCS-nodes** (IAMCCS) — VRAM Flush (soft) node, runs between clips in the batch loop to keep VRAM headroom stable across long ImageIterator runs. GPL-3.0. https://github.com/IAMCCS/IAMCCS-nodes

**Optional:**
• **ComfyUI-DaSiWa** (RTX GPU only) — DaSiWa_RTX_UpscalerRefiner for 1080p upscale branch. If missing (red node box), safe to ignore/delete — main path unaffected. https://github.com/darksidewalker/ComfyUI-DaSiWa-Nodes

Requires ComfyUI ≥ 0.30. H3 nodes (MiniMaxH3ImageToVideo, MiniMaxH3SigmaShift) are native to core.

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

🚀 **How to Use**

**Setup:**

1. Download all 5 model files → place in ComfyUI/models/ (exact paths above)
2. Install 4 custom node packs via ComfyUI Manager (Image_Anything via git clone if Manager fails)
3. Confirm ComfyUI ≥ 0.30
4. Load the workflow JSON into ComfyUI
5. On ImageIterator node: set folder_path to your input image directory

**Batch Generation:**

6. Set **Orientation INT** to 0 (landscape) or 1 (portrait) — matches input image orientation
7. Leave **Motion Switch = 0** (auto-caption) for simplest path; flip to 1 to type your own motion text
8. Click Queue → H3 loads input image → QwenVL auto-captions motion → cleanGpuUsed (VRAM bridge) → H3 samples 8 steps (turbo) → VAE decodes video + audio → CreateVideo produces MP4 with stereo audio
9. ImageIterator auto-advances to next image in folder on each Queue press

To batch 10 images: press Queue 10 times, or use Queue (Instant) ×10 from the UI.

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

⚙️ **Settings & Parameters**

| Setting | Default | Purpose |
|---------|---------|---------|
| **Orientation INT** | 0 | 0 = Landscape 1280×704; 1 = Portrait 704×1280 |
| **Motion Switch** | 0 | 0 = Auto (QwenVL captions frame); 1 = Manual (type MOTION TEXT) |
| **H3 Sampler** | res_multistep | Residual multistep (recommended for video) |
| **H3 Steps** | 6 | Dual-LoRA turbo stack range 4–8 tested, 6 = best speed/quality; non-turbo baseline is 20 |
| **H3 Scheduler** | simple | Simple scheduler (stable) |
| **Sigma Shift (video)** | 12.0 | Production-tuned; do not change |
| **Sigma Shift (audio)** | 3.0 | Audio emphasis; do not change |
| **Frames** | 243 | ~10.1s @ 24fps. Valid: 124/141/158/175/192/209/226/243 |
| **Seed** | 694207699552158 (randomize) | Ships randomize by default; set fixed for reproducible A/B testing |
| **Audio Input** | connected | Unplug wire to mute output |

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

💡 **Performance Tips & Troubleshooting**

**Minimum Spec** — 16 GB VRAM + ~32 GB system RAM (50GB+ if attempting frames ≥175).

**ImageIterator Folder Path** — use absolute paths (C:\path\to\images or /path/to/images); relative paths may fail.

**Red DaSiWa Node** — if ComfyUI-DaSiWa pack is not installed, the optional upscale side-branch shows red. This is safe — just bypass or delete the node. Main video path runs independently.

**Manager Can't Find Image_Anything** — ComfyUI Manager may not auto-resolve this custom node. Use git clone into custom_nodes/ instead (see Custom Nodes section).

**Frame Count & RAM** — 124–158 frames: ~32 GB RAM comfortable. ≥175 frames: ~48 GB system RAM needed (VRAM and system are separate). H3 frame grid is 17k+5 (not the LTX N*8+1 table).

**Turbo LoRA vs Non-Turbo** — this workflow ships with the dual-LoRA stack at 6 steps. To revert to the original 20-step path: bypass both LoRA nodes and raise BasicScheduler steps back to a non-turbo count.

**Motion Auto-Caption Quality** — QwenVL reads the input frame, so clearer, well-lit images produce richer motion captions. Abstract or low-contrast frames may yield generic descriptions (e.g., "slow ambient movement").

**Cold Start** — first workflow run: Qwen3-VL-2B-Instruct (~2.5 GB) auto-downloads (~1–2 min). Subsequent runs are fast (model cached).

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

📝 **Notes & AI Disclosure**

• **AI-Generated Content** — All example outputs are AI-generated by MiniMax H3. Respect local AI disclosure laws when publishing.
• **Configuration Only** — no model weights in JSON; download all 5 files separately from HuggingFace links above.
• **Workflow Reuse** — free to modify, share, and fork. No attribution needed (workflow is CC0).
• **Hardware Tested** — RTX 5080 16 GB VRAM + 48 GB system RAM (int8_convrot, native resolution, no upscale).

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

🔗 **Also Check Out**

Sister workflows by the same author:
• **Krea2 Turbo Dual** — still-image generator (text or reference image → image)
• **Krea2 → H3 Motion Suite** — text-to-image → video (SUBJECT + POSE → still → audio-video, one graph)

GitHub mirror: [github.com/thinnaphat-border/comfyui-workflows](https://github.com/thinnaphat-border/comfyui-workflows)

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

⭐ **Found this useful?**

• Like if it saved you time animating image batches
• Comment your results — I read every one
• Follow for new ComfyUI workflows, all tested on 16 GB VRAM

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

🎨 **Generate Your Input Images**

To create specialized source images for batch animation, try generating with Krea2 using these compatible LoRAs:
• **[Bokeh Dial](https://civitai.com/models/2815209)** — Depth-of-field control for cinematic focus effects
• **[y2ksnap](https://civitai.com/models/2802200)** — 2000s nostalgic aesthetic and styling
• **[mossworld](https://civitai.com/models/2812161)** — Overgrown Moss & Vines World Morph for nature textures

All three are Krea2-base compatible.

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

⚖️ **Model Attribution & Licensing**

**MiniMax H3** (Abiray int8_convrot quant)
• License: MiniMax H3 Community License — https://huggingface.co/Abiray/Minimax-H3-nvfp4-INT4-INT8-Convrot
• Commercial use ✅ OK; outputs are yours to sell/distribute
• The license's "Excluded Territories" clause restricts where the *model may be run* — check the license text for the current list before running. Output distribution is not territory-restricted.

**MiniMax H3 Turbo LoRA stack**
• Official Comfy-Org 8-step distill LoRA — Apache-2.0 — https://huggingface.co/Comfy-Org
• lightx2v turbo LoRA (secondary, str 0.75) — check upstream repo for current license before commercial redistribution

**Supporting Components**
• Qwen3-VL-32B-Instruct (H3 text encoder) & Qwen3-VL-2B-Instruct (motion caption VLM) — Apache-2.0
• H3 video/audio VAEs — MiniMax H3 Community License (part of the H3 release)

**ComfyUI Custom Nodes**
• ComfyUI-QwenVL (1038lab) — Apache-2.0/BSD — https://github.com/1038lab/ComfyUI-QwenVL
• ComfyUI_Image_Anything (Kelin) — https://github.com/ComfyUI-Kelin/ComfyUI_Image_Anything
• ComfyUI-Easy-Use (vjumpkung) — https://github.com/vjumpkung/ComfyUI-Easy-Use
• rgthree-comfy — Apache-2.0 — https://github.com/rgthree/rgthree-comfy
• IAMCCS-nodes — GPL-3.0 — https://github.com/IAMCCS/IAMCCS-nodes
• ComfyUI-DaSiWa (optional) — https://github.com/darksidewalker/ComfyUI-DaSiWa-Nodes

**Workflow JSON** — CC0 Public Domain. Free to use, modify, redistribute without attribution (credit appreciated).

All example outputs are AI-generated. Model weights remain the property of their respective owners and are downloaded separately from official HuggingFace sources.