P-H3-Image: Text to Image

Pruna AI's few-step image model on MiniMax-H3: a local prompt upsampler expands a plain request, then an 8-step sample renders it at your chosen size.

Input

Plain-language description of the scene; the graph's prompt upsampler expands it before sampling.

Shape of the image; 16:9 at 4 MP is about 2720 x 1536.

Pixel budget; sides round to a multiple of 32. 4 MP at 16:9 gives 2720 x 1536.

Advanced settings

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

Output
Output
Example

Try an example

About this workflow

Pruna AI's few-step image model built on MiniMax-H3: write a plain-language request and a local prompt upsampler expands it into a detailed scene before an 8-step sample. Pick the aspect ratio and megapixel target on the Resolution Selector to move between square, 4:3, 16:9 and portrait canvases.

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 prompt text and a ComfySwitchNode picks between the raw text and TextGenerateV2 (max_length 1536, thinking on, default template); PreviewAny shows the upsampled prompt. PH3ImageConditioning takes clip + prompt + vae, and width/height come from ResolutionSelector (4:3 (Standard), 4 MP, rounding multiple 32, which lands on 3360x2496). MiniMaxH3SigmaShift keeps shift_video and shift_audio at 1.0; BasicGuider -> KSamplerSelect (er_sde) with RandomNoise feed SamplerCustomAdvanced driven by 9 ManualSigmas (8 steps), then VAEDecode and SaveImage.

Inputs: the prompt, written as `Aspect ratio: <ratio>` then `Request: <what you want>`; aspect ratio, megapixels and rounding multiple on the Resolution Selector. Output: one image. Needs ComfyUI 0.44.0 and the p-h3-image files (int8 convrot checkpoints plus the bf16 LoRA).