MiniMax-H3-Image: Image Edit

Edit a picture by describing the change; MiniMax-H3-Image reads the source as <Picture 1> and rewrites only what you ask for.

Input

The picture to edit. The output follows its shape.

Say what to change and what to keep.

Advanced settings

Leave empty for a new result each run. Enter a number to repeat one.

Output

Your output will appear here.

关于此工作流

Edit a picture by describing the change. MiniMax-H3-Image reads the source as <Picture 1> and rewrites only what the instruction asks for, keeping the rest of the frame intact, so one checkpoint handles both text-to-image and reference-guided edits at about 16 MP. Ideal for swapping products, props or vehicles in a shot, changing a material, or restyling a still frame.

Graph: LoadImage (the shipped demo uses `vintage_yellow_taxi.png`) is scaled up to a 4 MP working budget and fed to MiniMaxH3ImageConditioning next to the caption. UNETLoader (`minimax_h3_image_pruned_int8_convrot.safetensors`), CLIPLoader (`qwen3vl_32b_minimax_h3_int8_convrot.safetensors`) and VAELoader load the open weights; a 6-step, CFG-1 KSampler samples and VAEDecode returns the picture. The prompt expander chain (TextGenerateV2, JsonExtractString, Enable PE toggle) rewrites the instruction before sampling. The output size follows the source picture, so this record exposes no size field. ImageCompare shows source and edit side by side.