Make a 16 megapixel picture from a prompt, or edit a picture in plain words. For an edit, you can give MiniMax H3 Image up to seven more pictures and tell it what to take from each one.
🎬▶️ Video: see this workflow in use ◀️🎬
How to use it
Drop the workflow into ComfyUI and press Run. It opens with a full prompt, so you get a picture right away.
For your own picture, write a full description. Or turn Auto prompt on and write a few words: a writer turns them into the long description the model wants.
To edit, turn Edit a picture on, load your picture and write the change, for example “Make it snow.” Auto prompt works here too: it looks at your pictures and writes out the change from a few words. Edit Lock keeps the parts you did not ask to change exactly where they were.
For a reference, connect an image loader to picture 2 on Write the prompt. Add more in order through picture 8. Name the picture in your request, for example “Put the jacket from picture 2 on the woman in picture 1.”
Size picks 4 MP, 8 MP or 16 MP. Fast runs 2 steps instead of 6. Grain cleanup clears fine speckle on skin, fur and stone: Light changes the picture least, Strong is the smoothest. Upscale the result doubles a 4 MP picture with SeedVR2.
What you need
Update ComfyUI and AusBoss nodes first. You need ComfyUI 0.45 or newer and AusBoss nodes 2.26.3 or newer. This version ran on ComfyUI 0.45.3 and AusBoss nodes 2.26.3. The workflow opens on High VRAM.
High VRAM · minimax_h3_image_pruned_int8_convrot.safetensors →
models/diffusion_models· 21.0 GBLow VRAM · minimax_h3_image_pruned_w4a8.safetensors →
models/diffusion_models· 12.5 GBBoth · qwen3vl_32b_minimax_h3_nvfp4_awq.safetensors →
models/text_encoders· 15.7 GBBoth · minimax_h3_image_vae_fp16.safetensors →
models/vae· 1.2 GBAuto prompt · qwen3.5_9b_w4a8_gptq.safetensors →
models/text_encoders· 7.3 GBFast · minimax_h3_pdmd_2nfe_comfyui.safetensors →
models/loras· 1.96 GB2× upscale · seedvr2_3b_int8_convrot.safetensors →
models/diffusion_models· 3.46 GB2× upscale · seedvr2_ema_vae_fp16.safetensors →
models/vae· 501 MB
Tips
On an RTX 5090 a new 16 MP picture took about 22 seconds and a 16 MP edit about 35. At 4 MP an edit took about 12.
Put any words that should appear in the picture in quotes. Signs and posters spell them right. Small text far from the subject can still come out as nonsense.
Leave Auto prompt off when your prompt is already a full description. In our runs of the opening prompt, rewriting it tangled the hands.
On an edit, Auto prompt makes a bigger change from a few words, like “vampire” or “winter”. For a small, exact change, leave it off.
For an edit, say what should stay the same as well as what you want to change.
If part of your edit is missing with Edit Lock on, switch it off and run again. The What Edit Lock kept readout says how much of your picture it held.
Grain cleanup also changes small details a little, like another try at the same picture. Fast skips it.
Fast frames a little tighter and small words can come out wrong. Leave it off for posters and covers.
MiniMax-H3-Image is by MiniMax and uses the MiniMax-H3-Image Community License. Powered by MiniMax-H3-Image. The 2-step LoRA is akhaliq's pDMD. Optional upscaling uses SeedVR2 under Apache 2.0.
v1.0 (2026-11-10) · First release.
v1.1 (2026-11-11) · Auto prompt now works on edits and sees every picture you plug in.
v1.2 (2026-11-20) · Edit Lock keeps the parts an edit did not change. Grain cleanup clears fine speckle. Size has 8 MP.
Report an issue · Discord · YouTube @ausbossdev · X @Zanzibased · GitHub
Description
Two things since v1.1. Edit Lock: on an edit, every part of your picture you did not ask to change stays exactly where it was. It needs no mask and has its own switch, on as it opens. When the edit changes the whole picture (a new style, new light, rain) it steps aside by itself, and a readout says how much of your picture it kept. Grain cleanup: a new Off | Light | Medium | Strong choice that clears the fine speckle a picture can get on skin, fur and stone. It takes no extra time and opens on Off, so pictures come out as before until you turn it on. Size has a new 8 MP choice between 4 MP and 16 MP, for a quicker edit on a smaller card. The details on each switch moved into a Settings + troubleshooting card you open at the top left. The layout is new too: the settings have their own group, and the groups are colored by job: teal for what you set, gray for what you can leave alone, green for the result. Needs ComfyUI 0.45 or newer and AusBoss nodes 2.26.3 or newer. Tested with ComfyUI 0.45.3 and AusBoss nodes 2.26.3.
