P-H3-Image: Image Edit

Pruna AI's few-step image model on MiniMax-H3: edit a source image from a short instruction with 8-step sampling, output keeps the source shape.

Input

The picture to edit. The result keeps its framing and pixel shape.

Say what to change and what to keep. The graph's prompt upsampler expands it, with the source as its vision input.

Advanced settings

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

Output
Output
Example

例を試す

このワークフローについて

Pruna AI's few-step image model built on MiniMax-H3: edit the source image with a short instruction, up to 9 reference images, 8-step sampling.

Graph: UNETLoader (minimax_h3_fl2va_pruned_int8_convrot.safetensors) into LoRA p-h3-image_lora_bf16 (strength 1.0); CLIPLoader with the p-h3-image int8 text encoder (type minimax) and a second CLIPLoader for the prompt upsampler (p-h3-image_prompt_upsampler_w4a8); VAELoader p-h3-image_vae_int8_convrot. A PrimitiveStringMultiline holds the instruction and a ComfySwitchNode picks between the raw text and TextGenerateV2 (max_length 1536, thinking on); PreviewAny shows the upsampled prompt. PH3ImageConditioning takes clip, prompt and vae plus images.image_1 from LoadImage; width and height are read from GetImageSize, so the result keeps the shape of the source. MiniMaxH3SigmaShift (1.0 / 1.0) -> BasicGuider -> KSamplerSelect er_sde with RandomNoise -> SamplerCustomAdvanced with 9 ManualSigmas (8 steps) -> VAEDecode -> SaveImage. ImageCompare shows source against result.

Inputs: one source image (LoadImage) and a prompt written as `Aspect ratio: <ratio>` then `Request: <the change>`; up to 9 reference images are accepted by the conditioning node. The upsampler's vision path (ImageScaleToTotalPixels, lanczos, 0.95 MP, multiple 32) is bypassed by default. Output: one edited image at the source dimensions.