DeepFloyd IF¶
| Model | Source | Revision | Weight license | Parity (CPU, fp32) |
|---|---|---|---|---|
deepfloyd-if-i-xl |
DeepFloyd/IF-I-XL-v1.0 | c03d510 |
DeepFloyd IF License (non-commercial) | Verified |
Catalog: deepfloyd-if ·
Models: deepfloyd-if-i-xl · Task: t2i
Architecture. A T5-XXL encoder conditions stage I pixel diffusion at 64×64. Stage II receives the same prompt context and the noisy bilinearly-upscaled stage-I image, then produces a 256×256 image. Both stages use learned-range DDPM variance and dynamic thresholding.
Usage.
from imagehub import pipeline
with pipeline("t2i", backend="deepfloyd-if-i-xl", model="checkpoints/deepfloyd-if-i-xl") as pipe:
result = pipe("a cat, oil painting", seed=42)
result.images[0].save("cat.png")
Settings. Output is fixed at 256×256. steps and superresolution_steps default to 50;
guidance and superresolution_guidance default to 7.0 and 4.0. noise_level defaults to 250.
negative_prompt, seed, and stage-I latents are also supported. Stage III, caption cleaning,
safety checking, watermarking and fp16/bfloat16 parity are out of scope.
License. The gated weights use the non-commercial DeepFloyd IF License. imagehub does not distribute weights.