MiniMax-H3-Image: Text to Image

Generate a photoreal or graphic image from a text prompt with MiniMax-H3-Image, an open-weights model distilled to six steps with no CFG.

Input

Describe the scene; the graph's prompt expander rewrites it before sampling.

Shape of the image; 1:1 at 4 MP is 2048 x 2048.

Pixel budget; sides round to a multiple of 32. The template starts at 4 MP.

Advanced settings

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

Output

Your output will appear here.

About this workflow

Generate a photoreal or graphic image from a text prompt with MiniMax-H3-Image, an open-weights model whose single checkpoint handles both text-to-image and image editing. Generation is distilled to 6 steps with no CFG and outputs about 16 MP natively, and an optional prompt expander turns a short request into the detailed caption the model reads best.

Graph: a subgraph holds UNETLoader (`minimax_h3_image_pruned_int8_convrot.safetensors`), CLIPLoader (`qwen3vl_32b_minimax_h3_int8_convrot.safetensors`), VAELoader (`minimax_h3_image_vae_fp16.safetensors`), MiniMaxH3ImageConditioning, VAEDecode and a 6-step KSampler at CFG 1 with the euler sampler. The prompt expander is a second CLIPLoader (`qwen3.5_9b_w4a8_gptq.safetensors`) driving TextGenerateV2, which rewrites the request into a caption that JsonExtractString pulls back out; the Enable PE toggle bypasses it. ResolutionSelector sets aspect ratio and megapixels, sides rounded to a multiple of 32 — the shipped demo is 1:1 at 4 MP — and SaveImage writes the result.