Models¶
123 checkpoints in 53 families, all behind pipeline("t2i", backend=<model id>, ...).
Each family page has the architecture notes, settings, license and the recorded parity result.
Stable Diffusion 3 and 3.5
Text→ImageParity · CPU fp32
Stability AI Non-Commercial Research Community License
Stable Diffusion 3 and 3.5
Text→ImageParity · CPU fp32
Stability AI Community License
PixelDiT
Text→ImageParity · CPU fp32
NVIDIA License (NSCLv1; non-commercial research or evaluation)
Image input¶
Most models take a text prompt only. The models in this table also accept an input image through the
image option (a PIL.Image.Image or an imagehub.Image); every model not listed here takes no
image input, and passing image to one raises an error.
| Model | Image input | Use | Options |
|---|---|---|---|
flux1-kontext-dev |
optional | instruction editing | image |
visualcloze-384 |
required | grid visual in-context generation | image |
joyai-image-edit |
optional | instruction editing | image |
joyai-image-edit-plus |
required | multi-image instruction editing | image |
qwen-image-edit |
required | instruction editing | image |
qwen-image-edit-2509 |
required | multi-image instruction editing | image |
sdxl-refiner-1.0 |
required | refinement (img2img) | image, strength |
omnigen-v1 |
optional | editing and subject-driven generation, referenced from the prompt | image, image_guidance, max_input_image_size, use_input_image_size_as_output |
LoRA adapters¶
LoRA adapter files load on top of a checkpoint in the Stable Diffusion 1.x and 2.x, SDXL, SD3, FLUX.1, FLUX.2 klein, Sana, Qwen-Image, Z-Image, Lumina 2, HiDream-I1, Krea 2 and Ideogram 4 backends, including the fine-tunes they serve. The U-Net backends patch the U-Net and the CLIP text encoders; the transformer backends patch the transformer.
from imagehub import pipeline
with pipeline("t2i", backend="sd-v1-5", model="checkpoints/sd-v1-5",
lora="style.safetensors", lora_scale=0.8) as pipe:
result = pipe("a watercolor fox", seed=0)
pipeline("t2i", backend="sdxl-base-1.0", model="checkpoints/sdxl-base-1.0",
lora=[("style.safetensors", 0.8), ("character.safetensors", 0.5)])
pipeline("t2i", backend="sd-v1-5", model="checkpoints/sd-v1-5",
lora="style.safetensors", lora_fuse=True)
The second call stacks two adapters with their own scales; the third merges the adapter into the weights after loading.
| Option | Meaning |
|---|---|
lora |
a .safetensors path, or a list of paths and (path, scale) pairs |
lora_scale |
the scale of a single path (default 1.0); use pairs for a list |
lora_fuse |
merge the adapters into the base weights after loading (default False) |
Layouts. The diffusers / PEFT layout is read on every backend above. The U-Net backends also read the legacy diffusers and kohya layouts:
| Layout | Example key |
|---|---|
| diffusers / PEFT | unet.down_blocks.0.attentions.0.transformer_blocks.0.attn1.to_q.lora_A.weight |
| legacy diffusers | unet.down_blocks.0.attentions.0.transformer_blocks.0.attn1.processor.to_q_lora.down.weight |
| kohya | lora_unet_down_blocks_0_attentions_0_proj_in.lora_down.weight (with .alpha) |
kohya SDXL files that use the original U-Net block names (input_blocks, middle_block,
output_blocks) are mapped to the diffusers names. Text-encoder keys are accepted with or without
CLIP's text_model level. Adapters can patch linear and convolution layers. The alpha comes from
the file (kohya .alpha tensors or the diffusers lora_adapter_metadata header) and equals the
rank when the file has none.
Rejected files. DoRA and LyCORIS (LoHa, LoKr) adapters, keys that match no module, and modules without both weights raise an error before the model is changed. Options are checked when the pipeline is created, before any weights load.
Programmatic use. imagehub.lora holds the layers (LoRALinear, LoRAConv2d) and the
inject_lora, merge_lora, unmerge_lora, set_lora_scale and remove_lora helpers;
imagehub.lora_formats.load_lora_file loads a file into any set of native components.
Outputs match diffusers load_lora_weights exactly on CPU in float32; the comparisons are in
the validation record.