support multi-stage prompts in prompt filter

api/onnx_web/chain/text_prompt.py

@@ -1,9 +1,11 @@
 from logging import getLogger
 from random import randint
+from re import sub
 from typing import Optional
 
 from transformers import pipeline
 
+from ..diffusers.utils import split_prompt
 from ..params import ImageParams, SizeChart, StageParams
 from ..server import ServerContext
 from ..worker import ProgressCallback, WorkerContext
@@ -13,6 +15,10 @@ from .result import StageResult
 logger = getLogger(__name__)
 
 
+LENGTH_MARGIN = 15
+RETRY_LIMIT = 5
+
+
 class TextPromptStage(BaseStage):
     max_tile = SizeChart.max
 
@@ -26,28 +32,55 @@ class TextPromptStage(BaseStage):
         *,
         callback: Optional[ProgressCallback] = None,
         prompt_model: str = "Gustavosta/MagicPrompt-Stable-Diffusion",
+        exclude_tokens: Optional[str] = None,
+        min_length: int = 75,
         **kwargs,
     ) -> StageResult:
         device = worker.device.torch_str()
-        gpt2_pipe = pipeline(
+        text_pipe = pipeline(
             "text-generation",
             model=prompt_model,
-            tokenizer="gpt2",
             device=device,
             framework="pt",
         )
 
-        input = params.prompt
-        max_length = len(input) + randint(60, 90)
-        logger.debug(
-            "generating new prompt with max length of %d from input prompt: %s",
-            max_length,
-            input,
-        )
+        prompt_parts = split_prompt(params.prompt)
+        prompt_results = []
+        for prompt in prompt_parts:
+            retries = 0
+            while len(prompt) < min_length and retries < RETRY_LIMIT:
+                max_length = len(prompt) + randint(
+                    min_length - LENGTH_MARGIN, min_length + LENGTH_MARGIN
+                )
+                logger.debug(
+                    "extending input prompt to max length of %d from %s: %s",
+                    max_length,
+                    len(prompt),
+                    prompt,
+                )
 
-        result = gpt2_pipe(input, max_length=max_length, num_return_sequences=1)
-        prompt = result[0]["generated_text"].strip()
-        logger.debug("replacing prompt with: %s", prompt)
+                result = text_pipe(
+                    prompt, max_length=max_length, num_return_sequences=1
+                )
+                prompt = result[0]["generated_text"].strip()
 
-        params.prompt = prompt
+                if exclude_tokens:
+                    logger.debug(
+                        "removing excluded tokens from prompt: %s", exclude_tokens
+                    )
+                    prompt = sub(exclude_tokens, "", prompt)
+
+            if retries >= RETRY_LIMIT:
+                logger.warning(
+                    "failed to extend input prompt to min length of %d, ended up with %d: %s",
+                    min_length,
+                    len(prompt),
+                    prompt,
+                )
+
+            prompt_results.append(prompt)
+
+        complete_prompt = " || ".join(prompt_results)
+        logger.debug("replacing input prompt: %s -> %s", params.prompt, complete_prompt)
+        params.prompt = complete_prompt
         return sources

api/onnx_web/diffusers/utils.py

@@ -475,9 +475,16 @@ def repair_nan(tile: np.ndarray) -> np.ndarray:
         return tile
 
 
+def split_prompt(prompt: str) -> List[str]:
+    if "||" in prompt:
+        return prompt.split("||")
+
+    return [prompt]
+
+
 def slice_prompt(prompt: str, slice: int) -> str:
     if "||" in prompt:
-        parts = prompt.split("||")
+        parts = split_prompt(prompt)
         return parts[min(slice, len(parts) - 1)]
     else:
         return prompt