sana: add complex human instruction to user prompts by default (untested)

helpers/caching/text_embeds.py

@@ -599,13 +599,18 @@ def compute_t5_prompt(self, prompt: str):
 
         return result, attn_mask
 
-    def compute_gemma_prompt(self, prompt: str):
+    def compute_gemma_prompt(self, prompt: str, is_negative_prompt: bool):
         prompt_embeds, prompt_attention_mask, _, _ = self.pipeline.encode_prompt(
             prompt=prompt,
             do_classifier_free_guidance=False,
             device=self.accelerator.device,
             clean_caption=False,
             max_sequence_length=300,
+            complex_human_instruction=(
+                StateTracker.get_args().sana_complex_human_instruction
+                if not is_negative_prompt
+                else None
+            ),
         )
 
         return prompt_embeds, prompt_attention_mask
@@ -616,6 +621,7 @@ def compute_embeddings_for_prompts(
         return_concat: bool = True,
         is_validation: bool = False,
         load_from_cache: bool = True,
+        is_negative_prompt: bool = False,
     ):
         logger.debug("Initialising text embed calculator...")
         if not self.batch_write_thread.is_alive():
@@ -694,6 +700,7 @@ def compute_embeddings_for_prompts(
                 raw_prompts,
                 return_concat=return_concat,
                 load_from_cache=load_from_cache,
+                is_negative_prompt=is_negative_prompt,
             )
         else:
             raise ValueError(
@@ -1073,6 +1080,7 @@ def compute_embeddings_for_sana_prompts(
         prompts: list = None,
         return_concat: bool = True,
         load_from_cache: bool = True,
+        is_negative_prompt: bool = False,
     ):
         logger.debug(
             f"compute_embeddings_for_sana_prompts arguments: prompts={prompts}, return_concat={return_concat}, load_from_cache={load_from_cache}"
@@ -1171,7 +1179,7 @@ def compute_embeddings_for_sana_prompts(
                             time.sleep(5)
                     # TODO: Batch this
                     prompt_embeds, attention_mask = self.compute_gemma_prompt(
-                        prompt=prompt,
+                        prompt=prompt, is_negative_prompt=is_negative_prompt
                     )
                     if "deepfloyd" not in StateTracker.get_args().model_type:
                         # we have to store the attn mask with the embed for pixart.

helpers/configuration/cmd_args.py

@@ -1469,7 +1469,24 @@ def get_argument_parser():
             " the value given."
         ),
     )
-
+    parser.add_argument(
+        "--sana_complex_human_instruction",
+        type=str,
+        default=[
+            "Given a user prompt, generate an 'Enhanced prompt' that provides detailed visual descriptions suitable for image generation. Evaluate the level of detail in the user prompt:",
+            "- If the prompt is simple, focus on adding specifics about colors, shapes, sizes, textures, and spatial relationships to create vivid and concrete scenes.",
+            "- If the prompt is already detailed, refine and enhance the existing details slightly without overcomplicating.",
+            "Here are examples of how to transform or refine prompts:",
+            "- User Prompt: A cat sleeping -> Enhanced: A small, fluffy white cat curled up in a round shape, sleeping peacefully on a warm sunny windowsill, surrounded by pots of blooming red flowers.",
+            "- User Prompt: A busy city street -> Enhanced: A bustling city street scene at dusk, featuring glowing street lamps, a diverse crowd of people in colorful clothing, and a double-decker bus passing by towering glass skyscrapers.",
+            "Please generate only the enhanced description for the prompt below and avoid including any additional commentary or evaluations:",
+            "User Prompt: ",
+        ],
+        help=(
+            "When generating embeds for Sana, a complex human instruction will be attached to your prompt by default."
+            " This is required for the Gemma model to produce meaningful image caption embeds."
+        ),
+    )
     parser.add_argument(
         "--allow_tf32",
         action="store_true",
@@ -2541,6 +2558,25 @@ def parse_cmdline_args(input_args=None, exit_on_error: bool = False):
             logger.error(f"Could not load skip layers: {e}")
             raise
 
+    if (
+        args.sana_complex_human_instruction is not None
+        and type(args.sana_complex_human_instruction) is str
+        and args.sana_complex_human_instruction not in ["", "None"]
+    ):
+        try:
+            import json
+
+            args.sana_complex_human_instruction = json.loads(
+                args.sana_complex_human_instruction
+            )
+        except Exception as e:
+            logger.error(
+                f"Could not load complex human instruction ({args.sana_complex_human_instruction}): {e}"
+            )
+            raise
+    elif args.sana_complex_human_instruction == "None":
+        args.sana_complex_human_instruction = None
+
     if args.enable_xformers_memory_efficient_attention:
         if args.attention_mechanism != "xformers":
             warning_log(

helpers/training/validation.py

@@ -258,11 +258,10 @@ def prepare_validation_prompt_list(args, embed_cache):
             or model_type == "sana"
         ):
             # we use the legacy encoder but we return no pooled embeds.
-            validation_negative_prompt_embeds = (
-                embed_cache.compute_embeddings_for_prompts(
-                    [StateTracker.get_args().validation_negative_prompt],
-                    load_from_cache=False,
-                )
+            validation_negative_prompt_embeds = embed_cache.compute_embeddings_for_prompts(
+                [StateTracker.get_args().validation_negative_prompt],
+                load_from_cache=False,
+                is_negative_prompt=True,  # sana needs this to disable Complex Human Instruction on negative embed generation
             )
 
             return (
@@ -1388,6 +1387,11 @@ def validate_prompt(
                 if StateTracker.get_model_family() == "flux":
                     if "negative_prompt" in pipeline_kwargs:
                         del pipeline_kwargs["negative_prompt"]
+                if self.args.model_family == "sana":
+                    pipeline_kwargs["complex_human_instruction"] = (
+                        self.args.sana_complex_human_instruction
+                    )
+
                 if (
                     StateTracker.get_model_family() == "pixart_sigma"
                     or StateTracker.get_model_family() == "smoldit"