Fix ErnieImagePipeline pre-computed prompt_embeds + num_images_per_prompt shape mismatch

[Ricardo-M-L]

What this PR does

When a user passes pre-computed prompt_embeds (or negative_prompt_embeds) alongside num_images_per_prompt > 1, ErnieImagePipeline.__call__ did not replicate the provided embeddings — the embeds list kept its original length (one per prompt) while the latents were allocated with total_batch_size = batch_size * num_images_per_prompt:

https://github.com/huggingface/diffusers/blob/main/src/diffusers/pipelines/ernie_image/pipeline_ernie_image.py#L287-L298

# [Phase 2] Text encoding
if prompt_embeds is not None:
    text_hiddens = prompt_embeds            # length = batch_size (NOT replicated)
else:
    text_hiddens = self.encode_prompt(prompt, device, num_images_per_prompt)

if self.do_classifier_free_guidance:
    if negative_prompt_embeds is not None:
        uncond_text_hiddens = negative_prompt_embeds   # same issue
    ...

Why this is a real bug

In the denoise loop, latents are expanded to total_batch_size:

https://github.com/huggingface/diffusers/blob/main/src/diffusers/pipelines/ernie_image/pipeline_ernie_image.py#L329-L343

if self.do_classifier_free_guidance:
    latent_model_input = torch.cat([latents, latents], dim=0)   # (batch*N*2, ...)

pred = self.transformer(
    hidden_states=latent_model_input,
    text_bth=text_bth,                                          # (batch*2, ...)
    ...
)

text_bth.shape[0] is derived from len(cfg_text_hiddens), which is len(prompt_embeds) * 2 under CFG — i.e. batch_size * 2, not batch_size * N * 2. This produces a shape mismatch inside the transformer's text-conditioning attention, and the call raises a RuntimeError. The standard "pre-compute embeds once, generate N variants" usage pattern is broken.

Minimal repro

pipe = ErnieImagePipeline.from_pretrained(...)
embeds = pipe.encode_prompt(["a cat"], pipe.device, num_images_per_prompt=1)
pipe(prompt_embeds=embeds, num_images_per_prompt=4, ...)
# RuntimeError during transformer forward

Fix

encode_prompt already performs this replication internally:

https://github.com/huggingface/diffusers/blob/main/src/diffusers/pipelines/ernie_image/pipeline_ernie_image.py#L158-L160

# Repeat for num_images_per_prompt
for _ in range(num_images_per_prompt):
    text_hiddens.append(hidden)

Mirror the same replication in the pre-embed branches so both paths yield a text_hiddens list of length batch_size * num_images_per_prompt:

 if prompt_embeds is not None:
-    text_hiddens = prompt_embeds
+    text_hiddens = [h for h in prompt_embeds for _ in range(num_images_per_prompt)]
 else:
     text_hiddens = self.encode_prompt(prompt, device, num_images_per_prompt)

 if self.do_classifier_free_guidance:
     if negative_prompt_embeds is not None:
-        uncond_text_hiddens = negative_prompt_embeds
+        uncond_text_hiddens = [h for h in negative_prompt_embeds for _ in range(num_images_per_prompt)]
     else:
         uncond_text_hiddens = self.encode_prompt(negative_prompt, device, num_images_per_prompt)

The non-embed path is unaffected since encode_prompt already replicates.

Before submitting

• Did you read the contributor guideline?
• Was this discussed/approved via a Github issue or the forum? N/A — standard API consistency fix.
• Did you make sure to update the documentation with your changes? N/A — no public API change.
• Did you write any new necessary tests? N/A — restores documented prompt_embeds + num_images_per_prompt combo.

Who can review?

@yiyixuxu @sayakpaul

[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.

[yiyixuxu]

thanks!
would you be willing to add a pipeline test for ernie-image? a new PR is ok