MiniMax H3 text-to-image and image editing in ComfyUI. Two starter workflows, one shared model setup, optional stackable LoRAs. Multiple reference images are supported.

The new path samples one latent frame and includes the video-VAE decode technique from Fizgig H3 Still. That latent is repeated only during decoding, then one image is kept. No frame counts, frame selection or separate Fizgig installation.

Still experimental: prompt adherence, identity, typography and fine detail vary. The decoder targets lone-frame banding; it cannot guarantee artifact-free images or successful edits.

Install and models

Use ComfyUI 0.37.0 or newer. Install Image Studio through ComfyUI Manager or clone this repository into ComfyUI/custom_nodes/. Restart ComfyUI and refresh the browser after updating.

Only three files are needed for the starting setup:

| Component | Starting file | Folder | |---|---|---| | H3 model | minimax_h3_fl2va_pruned_int8_convrot.safetensors | models/diffusion_models/ | | Text/vision encoder | qwen3vl_32b_minimax_h3_nvfp4_awq.safetensors | models/text_encoders/ | | Video VAE | minimax_h3_video_vae_int8_convrot.safetensors or your existing minimax_h3_video_vae_fp16.safetensors | models/vae/ |

Download from Comfy-Org/MiniMax-H3 using the official ComfyUI guide. Choose the quantization supported by your CUDA/PyTorch installation. Reuse files already on disk.

Following Fizgig's examples, both workflows start with FL2VA, including reference conditioning for editing. A separate REF2VA model is an optional comparison. No hybrid checkpoint, image VAE, audio VAE, Qwen refiner, decoder converter or model-download script is needed.

Two workflows

Open Templates → MiniMax H3 Image Studio, or drag either JSON onto ComfyUI:

Generate: Models → optional LoRAs → Text to Image → Render Image → Save Image.
Edit: add a Load Image connected to Image Edit.

Both templates include two H3 • LoRA nodes, bypassed by default. You can ignore them and generate without downloading any adapters.

Select the three files in H3 • Models. Write your prompt, choose aspect ratio and megapixels, then queue. Render Image exposes manual steps and seed. Changing only seed or steps reuses ComfyUI's cached prompt/reference preparation.

| Control | Starting point | |---|---| | Resolution | 3 MP; 1 MP for a preview, 4–8 MP for larger output | | Steps without a LoRA | 50, from Fizgig's no-Turbo still example; lower manually for speed | | Sampler / scheduler | ER-SDE / simple, under advanced controls | | Edit aspect ratio | source image, keeping its proportions at the chosen output area | | Exact dimensions | Select custom pixels, then set advanced width/height |

Sizes align to 32 pixels; MP uses ComfyUI's 1024² convention. There is no 0.98/1 MP restriction. Larger canvases cost sampling memory and time; tiled decoding does not make the diffusion model fit every GPU.

Editing and references

Image Edit combines image-to-image and reference editing. The source is <Picture 1>. Additional picture sockets grow as needed, up to nine pictures total. Each socket uses the first image of its input batch.

For a single source, write an ordinary instruction:

Replace the entire background with solid black. Keep the woman, her pose and clothing unchanged.

For multiple pictures, assign their roles explicitly:

Keep the person and clothes from <Picture 1>. Use the pose from <Picture 2>
and the room from <Picture 3>.

References are encoded separately from the generated still. There is no fixed frame-zero source anchor, hidden preservation-strength prompt or misleading denoise slider. This avoids returning a locked source frame, but the model can still miss an instruction.

Multiple LoRAs (optional)

Place compatible MiniMax H3 LoRAs in models/loras/. Each H3 • LoRA node has a filename and manual model strength. Enable the slots you want by switching their native node mode from Bypass to Always. Bypass unused slots; None or strength 0 also passes the models through without loading a LoRA.

Need more than two? Duplicate a LoRA node and connect it into the chain: Models → LoRA 1 → LoRA 2 → … → Text to Image / Image Edit. There is no hard-coded adapter-count limit. This works with generation, image-to-image and multi-reference editing; the text encoder and VAE remain shared. LoRAs for unrelated architectures are not supported. Each adapter may require a particular model, strength, sampler or step count—stacking them does not guarantee compatibility or better quality.

Optional faster recipe

For minimax_h3_turbo_v4_step600_ema.safetensors from Larry's H3 Turbo repository, enable one LoRA node, select the file, set strength 0.38 and Render steps 20, matching Fizgig's still example. All strengths and steps remain editable.

The older Models → advanced → turbo_lora shortcut still works at fixed 0.38 for saved workflows. Leave it None when applying that adapter through a LoRA node, so it is not applied twice. The community still recipe differs from Larry's strength-1, 4–8-step video guidance; an old LightX adapter is not interchangeable. Render's run_info lists the applied adapters and strengths.

Earlier versions

Reopen a new template after updating; saved canvases do not migrate automatically. Old Image Studio nodes remain registered as deprecated compatibility definitions and keep their original behavior.

Old workflows and measurements remain on GitHub in examples/ and the historical v23 guide. They are no longer starter examples or included in the Registry package. The Qwen detail refiner is also historical.

Feedback

Missing model choices: check folders, restart ComfyUI and refresh. Decoder errors: use the official H3 video VAE and current ComfyUI. Out of memory: lower megapixels or use supported quantized weights; native ComfyUI manages loading and offloading.

This project was coded with AI assistance. The author is not an experienced programmer; suggestions and bug reports are welcome. Include the traceback, workflow, versions, GPU/VRAM, model filenames, dimensions, steps and seed in a GitHub issue.

See validation, changelog and contributing. Image Studio code is Unlicense. The adapted Fizgig decoder keeps its MIT license and attribution in third-party notices. Models retain their own licenses.