Fix prepare latent image ids and vae sample generators for flux

[a-r-r-o-w]

No description provided.

[yiyixuxu]

thanks!

[HuggingFaceDocBuilderDev]

The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.

[a-r-r-o-w]

Wow, changing a randomly sampled latent (on every CI test run so far) to an actually properly seeded sampled latent breaks the test 😭

[yiyixuxu]

@a-r-r-o-w that's expected, no?

[a-r-r-o-w]

Oh sorry, I misunderstood/forgot what was going on. We did always seed the test before, but it was globally seeded. Now we use a different seed/generator passed into the pipeline so yes, this is expected.

[sayakpaul]

I can recollect two PRs here:

• [refactor] enhance readability of flux related pipelines #9711
• Flux latents fix #9929

Gently pinging @DN6

[a-r-r-o-w]

Oh, do you mean I am reverting a previous change that was necessary for something? I have been trying to pass latents for testing purposes but it always fails at this point unless the height and width are correctly adjusted -- it makes sense to me when comparing with other pipelines and doing rough calculation with tensor shapes, but LMK if this is wrong

[DN6]

No this is fine. We missed scaling the height and width when passing in latents. We do it here when latents aren't passed. Change is good 👍🏽

diffusers/src/diffusers/pipelines/flux/pipeline_flux.py

Line 528 in e564abe

latent_image_ids = self._prepare_latent_image_ids(batch_size, height // 2, width // 2, device, dtype)

[a-r-r-o-w]

Oh okay. Then I'll update the slices and merge once CI is green

[a-r-r-o-w]

I think only need to update slice for test_controlnet_flux.py despite there being similar changes for other flux pipelines with generator, because pipeline_flux_controlnet_image_to_image and pipeline_flux_controlnet_inpainting don't seem to have individual tests.