LLaDA-Image-Turbo-FP8

Text-to-image generation and single-image editing based on inclusionAI/LLaDA-Image-Turbo-FP8.

The latest version requires an access key in apikey (Secret) and a prompt. Provide input_image for single-image editing. aspect_ratio defaults to auto: editing follows the EXIF-corrected source ratio; text-to-image is square. Manual ratios center-crop the reference without stretching. Available ratios: auto, 1:1, 4:3, 3:4, 3:2, 2:3, 16:9, 9:16, 21:9, 9:21. Auto ratios must be between 1:4 and 4:1.

resolution_mode=longest_edge (default) produces a 1024px longest edge. Select megapixels mode for native 0.5–2.0MP output; the default area is 1.0MP. Requests above 2MP are rejected. Inference padding is cropped from the returned image. High resolutions use encoder CPU offload and take longer.

steps supports 2, 4 (default), and 6. Omit seed or pass null for randomness, or use 0–2147483647. file_format defaults to jpg (quality 90); png is lossless. Outputs are RGB.

Grouped FP8 expert kernels reduce warm text encoding to about 63 ms on L20, with default four-step generation around 2.96 seconds before saving. Hosted latency includes platform overhead. Higher-resolution edits can be softer and warmer; small text and precise typography may contain errors.

Version 947ea276170d970e6ad1575d51e2ca7aa1d83ac2c8fb7971832690d78a8d8cce uses this interface. Older immutable versions retain their original parameters.

Model revision: 664a975e4b4980fb750fd47b2f87e1874bb14bd7. Upstream code, revision 6bac7e7c1618e3ed9ec075ba75ff78cedfca6b38.