VisualCloze¶
| Model | Source | Revision | Weight license | Parity (CPU, fp32) |
|---|---|---|---|---|
visualcloze-384 (+ image input) |
VisualCloze/VisualClozePipeline-384 | 59c469d |
Apache-2.0 | Verified |
Models: visualcloze-384 · Task: t2i
VisualCloze generates the missing images in the final row of a visual in-context grid. Earlier
rows are complete examples; the final row supplies one or more condition images and uses None
for every target to generate. Results are returned left to right for those None entries.
from imagehub import pipeline
grid = [
[example_input, example_output],
[query_input, None],
]
with pipeline("t2i", backend="visualcloze-384", model="checkpoints/visualcloze-384") as pipe:
result = pipe(
"Each row shows a process ...",
image=grid,
content_prompt=None,
seed=0,
)
- Input.
imageis required. Every row has the same number of images, no earlier row may containNone, and the final row must have at least one.content_promptoptionally captions the target. The default settings are 30 steps, guidance 30.0, 512 text tokens and upsampling strength 0.3.height,width, negative prompts and batching options are unsupported. - Resolution. The grid-generation stage resizes each row around 384 pixels. Its targets are
then passed as uint8 images to a FLUX Fill upsampling stage. Set both
upsampling_heightandupsampling_width, or neither; the default uses the first supplied image in the final row, caps it at 1024² pixels and rounds down to multiples of 16. Setupsampling_strength=0to return the generation-stage targets directly. - Deliberate compatibility differences. imagehub uses the authors' plain layout instruction rather than the tuple representation emitted by diffusers, and keeps each target's native aspect during multi-target preprocessing rather than applying diffusers' swapped resize arguments.
- License. The published weights are Apache-2.0, but they are built on FLUX.1-Fill-dev, whose own license is non-commercial. Check the applicable terms before use.