[Bugfix] Fix image quality in /v1/images/generations for multi-stage pipeline (#2267)

Signed-off-by: Lancer <maruixiang6688@gmail.com>
Co-authored-by: Hongsheng Liu <liuhongsheng4@huawei.com>
This commit is contained in:
LancerandHongsheng Liu authored and GitHub committed 2026-04-17 20:09:41 +08:00
1 parent edb4f2fdcf
commit cf75ae6ed5
4 files changed
+559 -14

No files matched your search

+54 -2
View File
@@ -121,6 +121,7 @@ from vllm_omni.inputs.data import OmniDiffusionSamplingParams, OmniSamplingParam
logger = init_logger(__name__)
router = APIRouter()
MAX_UINT32_SEED = 2**32 - 1
profiler_router = APIRouter()
@@ -1312,6 +1313,57 @@ async def generate_images(request: ImageGenerationRequest, raw_request: Request)
)
try:
# Unify request construction for any multi-stage pipeline to avoid
# divergence between /v1/images and /v1/chat/completions.
if len(stage_configs) > 1:
chat_handler = getattr(raw_request.app.state, "openai_serving_chat", None)
if chat_handler is None:
logger.warning("openai_serving_chat is not initialized for multi-stage /v1/images/generations")
raise HTTPException(
status_code=HTTPStatus.SERVICE_UNAVAILABLE.value,
detail="openai_serving_chat is not initialized for multi-stage image generation.",
)
effective_seed = request.seed if request.seed is not None else random.randint(0, MAX_UINT32_SEED)
extra_body: dict[str, Any] = {
"seed": effective_seed,
"num_outputs_per_prompt": request.n,
}
if request.size is not None:
parse_size(request.size)
width, height = parse_size(request.size)
app_state_args = getattr(raw_request.app.state, "args", None)
_check_max_generated_image_size(app_state_args, width, height)
extra_body["size"] = request.size
if request.negative_prompt is not None:
extra_body["negative_prompt"] = request.negative_prompt
if request.num_inference_steps is not None:
extra_body["num_inference_steps"] = request.num_inference_steps
if request.guidance_scale is not None:
extra_body["guidance_scale"] = request.guidance_scale
if request.true_cfg_scale is not None:
extra_body["true_cfg_scale"] = request.true_cfg_scale
if request.generator_device is not None:
extra_body["generator_device"] = request.generator_device
if request.lora is not None:
# Keep /images validation semantics: invalid LoRA should fail with 400.
_parse_lora_request(request.lora)
extra_body["lora"] = request.lora
generation_result = await chat_handler.generate_diffusion_images(
prompt=request.prompt,
extra_body=extra_body,
request_id=f"img_gen-{random_uuid()}",
)
if isinstance(generation_result, ErrorResponse):
return JSONResponse(
status_code=generation_result.error.code if generation_result.error else 400,
content=generation_result.model_dump(),
)
flat_images, _, _ = generation_result
image_data = [ImageData(b64_json=encode_image_base64(img), revised_prompt=None) for img in flat_images]
return ImageGenerationResponse(created=int(time.time()), data=image_data)
# Build params - pass through user values directly
prompt: OmniTextPrompt = {"prompt": request.prompt}
if request.negative_prompt is not None:
@@ -1352,7 +1404,7 @@ async def generate_images(request: ImageGenerationRequest, raw_request: Request)
# This fixes issues where using the default global generator
# might produce blurry images in some environments.
_update_if_not_none(
gen_params, "seed", request.seed if request.seed is not None else random.randint(0, 2**32 - 1)
gen_params, "seed", request.seed if request.seed is not None else random.randint(0, MAX_UINT32_SEED)
)
_update_if_not_none(gen_params, "generator_device", request.generator_device)
_update_if_not_none(gen_params, "layers", request.layers)
@@ -1573,7 +1625,7 @@ async def edit_images(
# a proper generator is initialized in the backend.
# This fixes issues where using the default global generator
# might produce blurry images in some environments.
_update_if_not_none(gen_params, "seed", seed if seed is not None else random.randint(0, 2**32 - 1))
_update_if_not_none(gen_params, "seed", seed if seed is not None else random.randint(0, MAX_UINT32_SEED))
_update_if_not_none(gen_params, "generator_device", generator_device)
_update_if_not_none(gen_params, "layers", layers)
_update_if_not_none(gen_params, "resolution", resolution)