fix(flux): tighten check_inputs validation#13955
Conversation
…olnet img2img dims, redux scales
|
Hi @akshan-main, thanks for the PR! It does not appear to link an issue it fixes. If this PR addresses an existing issue, please add a closing keyword (e.g. |
| raise ValueError(f"The value of strength should in [0.0, 1.0] but is {strength}") | ||
|
|
||
| if height % self.vae_scale_factor * 2 != 0 or width % self.vae_scale_factor * 2 != 0: | ||
| if height % (self.vae_scale_factor * 2) != 0 or width % (self.vae_scale_factor * 2) != 0: |
There was a problem hiding this comment.
What value is this adding though?
There was a problem hiding this comment.
For consistency, all flux pipelines have the same 2 * (height // (vae_scale_factor * 2)) round-down in prepare_latents, but the others raise in check_inputs first, so an invalid size like 72 errors there and never reaches it. This is the only flux pipeline that silently resizes to 64 instead of erroring. Precedencies!
| if prompt_embeds is not None and negative_prompt_embeds is not None: | ||
| if prompt_embeds.shape != negative_prompt_embeds.shape: | ||
| raise ValueError( | ||
| "`prompt_embeds` and `negative_prompt_embeds` must have the same shape when passed directly, but" | ||
| f" got: `prompt_embeds` {prompt_embeds.shape} != `negative_prompt_embeds`" | ||
| f" {negative_prompt_embeds.shape}." | ||
| ) | ||
|
|
There was a problem hiding this comment.
I think this should also be checked when do_true_cfg is True? Otherwise, it's not used right?
There was a problem hiding this comment.
moved it into the do_true_cfg branch in both base flux and kontext, so the shape check only runs when the negative embeds are used.
github-actions
Bot
added
tests
size/M
| with self.assertRaises(ValueError): | ||
| pipe(**base_inputs, true_cfg_scale=2.0, generator=torch.manual_seed(0)) |
There was a problem hiding this comment.
Let's also inspect the message and assert around it.
|
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. |
3 participants
What does this PR do?
Addresses part of the flux review (#13584) - three input-validation gaps in the flux pipelines:
FluxPipeline/FluxKontextPipelineacceptednegative_prompt_embedswith a different shape thanprompt_embeds. Added the shape check the img2img/inpaint variants already have (propagated to kontext via# Copied from).FluxControlNetImg2ImgPipelinecheckedheight % self.vae_scale_factor * 2, which parses as(height % vae_scale_factor) * 2, so invalid sizes like 72 passed. Added the parentheses to match every other flux pipeline.FluxPriorReduxPipelineonly validated scale length for list images and ignoredpooled_prompt_embeds_scale. Now computes the image batch size for tensors too and validates both scales.Verified each finding's repro now raises/flags, and that matching/valid inputs still pass.
Before submitting
Who can review?
@hlky @sayakpaul @yiyixuxu