Ana içeriğe geç

Chroma

Model Source Revision Weight license Parity (CPU, fp32)
chroma1-hd lodestones/Chroma1-HD 0e0c60e Apache-2.0 Verified
chroma1-flash lodestones/Chroma1-Flash 093c4f6 Apache-2.0 Verified
chroma1-base lodestones/Chroma1-Base 50ebe12 Apache-2.0 Verified

Catalog: chroma1-hd, chroma1-flash, chroma1-base · Models: chroma1-hd, chroma1-flash, chroma1-base · Task: t2i

Architecture. Chroma1-HD is a ~8.9B diffusion transformer built on the FLUX.1 latent-flow foundation. It reuses the FLUX.1 VAE (16 latent channels, scaling factor 0.3611, shift factor 0.1159) and the flow-match Euler scheduler with exponential shifting. The text encoder is T5-XXL (4.7B parameters, bfloat16 weights) with a native SentencePiece Unigram tokenizer. The new Chroma transformer replaces FLUX.1's timestep-plus-pooled-text embedding with a distilled-guidance approximator: a sinusoidal timestep projection is combined with a fixed modulation projection and fed through a small MLP to produce per-block scale, shift and gate vectors for all 19 double-stream and 38 single-stream blocks. Attention is joint over text and image tokens, using 3-axis RoPE and an attention mask that keeps the real text tokens plus one padding token (the remaining padding tokens are masked out). Classifier-free guidance is applied when guidance > 1; otherwise the negative pass is skipped.

Usage.

python -m imagehub.pretrained chroma1-hd checkpoints/chroma1-hd
from imagehub import pipeline
with pipeline("t2i", backend="chroma1-hd", model="checkpoints/chroma1-hd", device="mps") as pipe:
    result = pipe("a cat, oil painting", seed=42)
    result.images[0].save("cat.png")

Settings. Chroma1-HD defaults to 35 steps and guidance 5.0; Chroma1-Flash defaults to 4 steps and guidance 1.0; Chroma1-Base defaults to 40 steps and guidance 3.0 and uses the checkpoint's Beta(0.6, 0.6) flow-match sigma schedule. All default to 1024×1024; width and height must be multiples of 16. The other supported options are num_images, seed, latents, negative_prompt and max_sequence_length (default 512, at most 512). Unrecognized settings raise an error.

Seed compatibility. Random-weight parity against diffusers 0.40.0 ChromaPipeline is exact on CPU in float32. The published bfloat16 checkpoints are loaded as float32 and reproduce the recorded reference outputs exactly (validation). MPS and CUDA runs are smoke tests only; there is no claim of exact equality on these devices.

Out of scope. Img2img, inpainting, IP-Adapter, LoRA, single-file checkpoint loading and fp16/bf16 parity are not supported.

License. The weights use the Apache-2.0 license. imagehub does not distribute weights.

Reference and imagehub, side by side

Drag the slider. Left is the recorded reference, right is imagehub, from the same seed, on CPU in float32.