feat(api): prefer chain stage parameters over request parameters (#138)

api/onnx_web/chain/blend_img2img.py

@@ -19,15 +19,13 @@ def blend_img2img(
     server: ServerContext,
     _stage: StageParams,
     params: ImageParams,
-    source_image: Image.Image,
+    source: Image.Image,
     *,
-    strength: float,
-    prompt: Optional[str] = None,
     callback: ProgressCallback = None,
     **kwargs,
 ) -> Image.Image:
-    prompt = prompt or params.prompt
-    logger.info("blending image using img2img, %s steps: %s", params.steps, prompt)
+    params = params.with_args(**kwargs)
+    logger.info("blending image using img2img, %s steps: %s", params.steps, params.prompt)
 
     pipe = load_pipeline(
         server,
@@ -41,25 +39,25 @@ def blend_img2img(
         logger.debug("using LPW pipeline for img2img")
         rng = torch.manual_seed(params.seed)
         result = pipe.img2img(
-            prompt,
+            params.prompt,
             generator=rng,
             guidance_scale=params.cfg,
-            image=source_image,
+            image=source,
             negative_prompt=params.negative_prompt,
             num_inference_steps=params.steps,
-            strength=strength,
+            strength=params.strength,
             callback=callback,
         )
     else:
         rng = np.random.RandomState(params.seed)
         result = pipe(
-            prompt,
+            params.prompt,
             generator=rng,
             guidance_scale=params.cfg,
-            image=source_image,
+            image=source,
             negative_prompt=params.negative_prompt,
             num_inference_steps=params.steps,
-            strength=strength,
+            strength=params.strength,
             callback=callback,
         )
 

api/onnx_web/chain/blend_inpaint.py

@@ -22,42 +22,44 @@ def blend_inpaint(
     server: ServerContext,
     stage: StageParams,
     params: ImageParams,
-    source_image: Image.Image,
+    source: Image.Image,
     *,
     expand: Border,
-    mask_image: Optional[Image.Image] = None,
+    mask: Optional[Image.Image] = None,
     fill_color: str = "white",
     mask_filter: Callable = mask_filter_none,
     noise_source: Callable = noise_source_histogram,
     callback: ProgressCallback = None,
     **kwargs,
 ) -> Image.Image:
+    params = params.with_args(**kwargs)
+    expand = expand.with_args(**kwargs)
     logger.info(
         "blending image using inpaint, %s steps: %s", params.steps, params.prompt
     )
 
-    if mask_image is None:
+    if mask is None:
         # if no mask was provided, keep the full source image
-        mask_image = Image.new("RGB", source_image.size, "black")
+        mask = Image.new("RGB", source.size, "black")
 
-    source_image, mask_image, noise_image, _full_dims = expand_image(
-        source_image,
-        mask_image,
+    source, mask, noise, _full_dims = expand_image(
+        source,
+        mask,
         expand,
         fill=fill_color,
         noise_source=noise_source,
         mask_filter=mask_filter,
     )
 
     if is_debug():
-        save_image(server, "last-source.png", source_image)
-        save_image(server, "last-mask.png", mask_image)
-        save_image(server, "last-noise.png", noise_image)
+        save_image(server, "last-source.png", source)
+        save_image(server, "last-mask.png", mask)
+        save_image(server, "last-noise.png", noise)
 
     def outpaint(image: Image.Image, dims: Tuple[int, int, int]):
         left, top, tile = dims
         size = Size(*image.size)
-        mask = mask_image.crop((left, top, left + tile, top + tile))
+        mask = mask.crop((left, top, left + tile, top + tile))
 
         if is_debug():
             save_image(server, "tile-source.png", image)
@@ -108,7 +110,7 @@ def outpaint(image: Image.Image, dims: Tuple[int, int, int]):
         return result.images[0]
 
     output = process_tile_order(
-        stage.tile_order, source_image, SizeChart.auto, 1, [outpaint]
+        stage.tile_order, source, SizeChart.auto, 1, [outpaint]
     )
 
     logger.info("final output image size", output.size)

api/onnx_web/chain/blend_mask.py

@@ -19,7 +19,7 @@ def blend_mask(
     _stage: StageParams,
     _params: ImageParams,
     *,
-    resized: Optional[List[Image.Image]] = None,
+    sources: Optional[List[Image.Image]] = None,
     mask: Optional[Image.Image] = None,
     _callback: ProgressCallback = None,
     **kwargs,
@@ -36,7 +36,7 @@ def blend_mask(
 
     resized = [
         valid_image(s, min_dims=mult_mask.size, max_dims=mult_mask.size)
-        for s in resized
+        for s in sources
     ]
 
     return Image.composite(resized[0], resized[1], mult_mask)

api/onnx_web/chain/correct_codeformer.py

@@ -25,6 +25,8 @@ def correct_codeformer(
     # must be within the load function for patch to take effect
     from codeformer import CodeFormer
 
+    upscale = upscale.with_args(**kwargs)
+
     device = job.get_device()
     pipe = CodeFormer(upscale=upscale.face_outscale).to(device.torch_str())
     return pipe(stage_source or source)

api/onnx_web/chain/correct_gfpgan.py

@@ -55,6 +55,8 @@ def correct_gfpgan(
     upscale: UpscaleParams,
     **kwargs,
 ) -> Image.Image:
+    upscale = upscale.with_args(**kwargs)
+
     if upscale.correction_model is None:
         logger.warn("no face model given, skipping")
         return source_image

api/onnx_web/chain/source_txt2img.py

@@ -21,12 +21,12 @@ def source_txt2img(
     source_image: Image.Image,
     *,
     size: Size,
-    prompt: str = None,
     callback: ProgressCallback = None,
     **kwargs,
 ) -> Image.Image:
-    prompt = prompt or params.prompt
-    logger.info("generating image using txt2img, %s steps: %s", params.steps, prompt)
+    params = params.with_args(**kwargs)
+    size = size.with_args(**kwargs)
+    logger.info("generating image using txt2img, %s steps: %s", params.steps, params.prompt)
 
     if source_image is not None:
         logger.warn(
@@ -47,7 +47,7 @@ def source_txt2img(
         logger.debug("using LPW pipeline for txt2img")
         rng = torch.manual_seed(params.seed)
         result = pipe.text2img(
-            prompt,
+            params.prompt,
             height=size.height,
             width=size.width,
             generator=rng,
@@ -60,7 +60,7 @@ def source_txt2img(
     else:
         rng = np.random.RandomState(params.seed)
         result = pipe(
-            prompt,
+            params.prompt,
             height=size.height,
             width=size.width,
             generator=rng,

api/onnx_web/chain/upscale_stable_diffusion.py

@@ -69,12 +69,12 @@ def upscale_stable_diffusion(
     source: Image.Image,
     *,
     upscale: UpscaleParams,
-    prompt: str = None,
     callback: ProgressCallback = None,
     **kwargs,
 ) -> Image.Image:
-    prompt = prompt or params.prompt
-    logger.info("upscaling with Stable Diffusion, %s steps: %s", params.steps, prompt)
+    params = params.with_args(**kwargs)
+    upscale = upscale.with_args(**kwargs)
+    logger.info("upscaling with Stable Diffusion, %s steps: %s", params.steps, params.prompt)
 
     pipeline = load_stable_diffusion(server, upscale, job.get_device())
     generator = torch.manual_seed(params.seed)

api/onnx_web/diffusion/run.py

@@ -255,7 +255,7 @@ def run_blend_pipeline(
         server,
         stage,
         params,
-        resized=sources,
+        sources=sources,
         mask=mask,
         callback=progress,
     )

api/onnx_web/params.py

@@ -47,6 +47,14 @@ def tojson(self):
             "bottom": self.bottom,
         }
 
+    def with_args(self, **kwargs):
+        return Border(
+            kwargs.get("left", self.left),
+            kwargs.get("right", self.right),
+            kwargs.get("top", self.top),
+            kwargs.get("bottom", self.bottom),
+        )
+
     @classmethod
     def even(cls, all: int):
         return Border(all, all, all, all)
@@ -75,6 +83,12 @@ def tojson(self) -> Dict[str, int]:
             "width": self.width,
         }
 
+    def with_args(self, **kwargs):
+        return Size(
+            kwargs.get("height", self.height),
+            kwargs.get("width", self.width),
+        )
+
 
 class DeviceParams:
     def __init__(
@@ -156,13 +170,25 @@ def tojson(self) -> Dict[str, Optional[Param]]:
             "model": self.model,
             "scheduler": self.scheduler.__name__,
             "prompt": self.prompt,
-            "negativePrompt": self.negative_prompt,
+            "negative_prompt": self.negative_prompt,
             "cfg": self.cfg,
             "seed": self.seed,
             "steps": self.steps,
             "lpw": self.lpw,
         }
 
+    def with_args(self, **kwargs):
+        return ImageParams(
+            kwargs.get("model", self.model),
+            kwargs.get("scheduler", self.scheduler),
+            kwargs.get("prompt", self.prompt),
+            kwargs.get("cfg", self.cfg),
+            kwargs.get("steps", self.steps),
+            kwargs.get("seed", self.seed),
+            kwargs.get("negative_prompt", self.negative_prompt),
+            kwargs.get("lpw", self.lpw),
+        )
+
 
 class StageParams:
     """
@@ -259,3 +285,20 @@ def tojson(self):
             "tile_pad": self.tile_pad,
             "upscale_order": self.upscale_order,
         }
+
+    def with_args(self, **kwargs):
+        return ImageParams(
+            kwargs.get("upscale_model", self.upscale_model),
+            kwargs.get("correction_model", self.correction_model),
+            kwargs.get("denoise", self.denoise),
+            kwargs.get("faces", self.faces),
+            kwargs.get("face_outscale", self.face_outscale),
+            kwargs.get("face_strength", self.face_strength),
+            kwargs.get("format", self.format),
+            kwargs.get("half", self.half),
+            kwargs.get("outscale", self.outscale),
+            kwargs.get("pre_pad", self.pre_pad),
+            kwargs.get("scale", self.scale),
+            kwargs.get("tile_pad", self.tile_pad),
+            kwargs.get("upscale_order", self.upscale_order),
+        )