diff options
| -rw-r--r-- | modules/infotext_utils.py | 3 | ||||
| -rw-r--r-- | modules/processing.py | 1 | ||||
| -rw-r--r-- | modules/sd_emphasis.py | 70 | ||||
| -rw-r--r-- | modules/sd_hijack_clip.py | 21 | ||||
| -rw-r--r-- | modules/sd_hijack_clip_old.py | 2 | ||||
| -rw-r--r-- | modules/shared_options.py | 5 |
6 files changed, 89 insertions, 13 deletions
diff --git a/modules/infotext_utils.py b/modules/infotext_utils.py index 1049c6c3..a938aa2a 100644 --- a/modules/infotext_utils.py +++ b/modules/infotext_utils.py @@ -356,6 +356,9 @@ Steps: 20, Sampler: Euler a, CFG scale: 7, Seed: 965400086, Size: 512x512, Model if "Cache FP16 weight for LoRA" not in res and res["FP8 weight"] != "Disable":
res["Cache FP16 weight for LoRA"] = False
+ if "Emphasis" not in res:
+ res["Emphasis"] = "Original"
+
infotext_versions.backcompat(res)
for key in skip_fields:
diff --git a/modules/processing.py b/modules/processing.py index 52f00bfb..f4aa165d 100644 --- a/modules/processing.py +++ b/modules/processing.py @@ -455,6 +455,7 @@ class StableDiffusionProcessing: self.height,
opts.fp8_storage,
opts.cache_fp16_weight,
+ opts.emphasis,
)
def get_conds_with_caching(self, function, required_prompts, steps, caches, extra_network_data, hires_steps=None):
diff --git a/modules/sd_emphasis.py b/modules/sd_emphasis.py new file mode 100644 index 00000000..654817b6 --- /dev/null +++ b/modules/sd_emphasis.py @@ -0,0 +1,70 @@ +from __future__ import annotations
+import torch
+
+
+class Emphasis:
+ """Emphasis class decides how to death with (emphasized:1.1) text in prompts"""
+
+ name: str = "Base"
+ description: str = ""
+
+ tokens: list[list[int]]
+ """tokens from the chunk of the prompt"""
+
+ multipliers: torch.Tensor
+ """tensor with multipliers, once for each token"""
+
+ z: torch.Tensor
+ """output of cond transformers network (CLIP)"""
+
+ def after_transformers(self):
+ """Called after cond transformers network has processed the chunk of the prompt; this function should modify self.z to apply the emphasis"""
+
+ pass
+
+
+class EmphasisNone(Emphasis):
+ name = "None"
+ description = "disable the mechanism entirely and treat (:.1.1) as literal characters"
+
+
+class EmphasisIgnore(Emphasis):
+ name = "Ignore"
+ description = "treat all empasised words as if they have no emphasis"
+
+
+class EmphasisOriginal(Emphasis):
+ name = "Original"
+ description = "the orginal emphasis implementation"
+
+ def after_transformers(self):
+ original_mean = self.z.mean()
+ self.z = self.z * self.multipliers.reshape(self.multipliers.shape + (1,)).expand(self.z.shape)
+
+ # restoring original mean is likely not correct, but it seems to work well to prevent artifacts that happen otherwise
+ new_mean = self.z.mean()
+ self.z = self.z * (original_mean / new_mean)
+
+
+class EmphasisOriginalNoNorm(EmphasisOriginal):
+ name = "No norm"
+ description = "same as orginal, but without normalization (seems to work better for SDXL)"
+
+ def after_transformers(self):
+ self.z = self.z * self.multipliers.reshape(self.multipliers.shape + (1,)).expand(self.z.shape)
+
+
+def get_current_option(emphasis_option_name):
+ return next(iter([x for x in options if x.name == emphasis_option_name]), EmphasisOriginal)
+
+
+def get_options_descriptions():
+ return ", ".join(f"{x.name}: {x.description}" for x in options)
+
+
+options = [
+ EmphasisNone,
+ EmphasisIgnore,
+ EmphasisOriginal,
+ EmphasisOriginalNoNorm,
+]
diff --git a/modules/sd_hijack_clip.py b/modules/sd_hijack_clip.py index 89634fbf..98350ac4 100644 --- a/modules/sd_hijack_clip.py +++ b/modules/sd_hijack_clip.py @@ -3,7 +3,7 @@ from collections import namedtuple import torch
-from modules import prompt_parser, devices, sd_hijack
+from modules import prompt_parser, devices, sd_hijack, sd_emphasis
from modules.shared import opts
@@ -88,7 +88,7 @@ class FrozenCLIPEmbedderWithCustomWordsBase(torch.nn.Module): Returns the list and the total number of tokens in the prompt.
"""
- if opts.enable_emphasis:
+ if opts.emphasis != "None":
parsed = prompt_parser.parse_prompt_attention(line)
else:
parsed = [[line, 1.0]]
@@ -249,6 +249,9 @@ class FrozenCLIPEmbedderWithCustomWordsBase(torch.nn.Module): hashes.append(self.hijack.extra_generation_params.get("TI hashes"))
self.hijack.extra_generation_params["TI hashes"] = ", ".join(hashes)
+ if any(x for x in texts if "(" in x or "[" in x) and opts.emphasis != "Original":
+ self.hijack.extra_generation_params["Emphasis"] = opts.emphasis
+
if getattr(self.wrapped, 'return_pooled', False):
return torch.hstack(zs), zs[0].pooled
else:
@@ -274,14 +277,14 @@ class FrozenCLIPEmbedderWithCustomWordsBase(torch.nn.Module): pooled = getattr(z, 'pooled', None)
- # restoring original mean is likely not correct, but it seems to work well to prevent artifacts that happen otherwise
- batch_multipliers = torch.asarray(batch_multipliers).to(devices.device)
- original_mean = z.mean()
- z = z * batch_multipliers.reshape(batch_multipliers.shape + (1,)).expand(z.shape)
- new_mean = z.mean()
+ emphasis = sd_emphasis.get_current_option(opts.emphasis)()
+ emphasis.tokens = remade_batch_tokens
+ emphasis.multipliers = torch.asarray(batch_multipliers).to(devices.device)
+ emphasis.z = z
+
+ emphasis.after_transformers()
- if not getattr(opts, "disable_normalize_embeddings", False):
- z = z * (original_mean / new_mean)
+ z = emphasis.z
if pooled is not None:
z.pooled = pooled
diff --git a/modules/sd_hijack_clip_old.py b/modules/sd_hijack_clip_old.py index c5c6270b..43e9b952 100644 --- a/modules/sd_hijack_clip_old.py +++ b/modules/sd_hijack_clip_old.py @@ -32,7 +32,7 @@ def process_text_old(self: sd_hijack_clip.FrozenCLIPEmbedderWithCustomWordsBase, embedding, embedding_length_in_tokens = self.hijack.embedding_db.find_embedding_at_position(tokens, i)
- mult_change = self.token_mults.get(token) if shared.opts.enable_emphasis else None
+ mult_change = self.token_mults.get(token) if shared.opts.emphasis != "None" else None
if mult_change is not None:
mult *= mult_change
i += 1
diff --git a/modules/shared_options.py b/modules/shared_options.py index 417a42b2..ba6d731d 100644 --- a/modules/shared_options.py +++ b/modules/shared_options.py @@ -1,7 +1,7 @@ import os
import gradio as gr
-from modules import localization, ui_components, shared_items, shared, interrogate, shared_gradio_themes, util
+from modules import localization, ui_components, shared_items, shared, interrogate, shared_gradio_themes, util, sd_emphasis
from modules.paths_internal import models_path, script_path, data_path, sd_configs_path, sd_default_config, sd_model_file, default_sd_model_file, extensions_dir, extensions_builtin_dir, default_output_dir # noqa: F401
from modules.shared_cmd_options import cmd_opts
from modules.options import options_section, OptionInfo, OptionHTML, categories
@@ -154,8 +154,7 @@ options_templates.update(options_section(('sd', "Stable Diffusion", "sd"), { "sd_checkpoint_cache": OptionInfo(0, "Checkpoints to cache in RAM", gr.Slider, {"minimum": 0, "maximum": 10, "step": 1}).info("obsolete; set to 0 and use the two settings above instead"),
"sd_unet": OptionInfo("Automatic", "SD Unet", gr.Dropdown, lambda: {"choices": shared_items.sd_unet_items()}, refresh=shared_items.refresh_unet_list).info("choose Unet model: Automatic = use one with same filename as checkpoint; None = use Unet from checkpoint"),
"enable_quantization": OptionInfo(False, "Enable quantization in K samplers for sharper and cleaner results. This may change existing seeds").needs_reload_ui(),
- "enable_emphasis": OptionInfo(True, "Enable emphasis").info("use (text) to make model pay more attention to text and [text] to make it pay less attention"),
- "disable_normalize_embeddings": OptionInfo(False, "Disable normalize embeddings").info("Do not normalize embeddings after calculating emphasis. It can be expected to be effective in preventing artifacts in SDXL."),
+ "emphasis": OptionInfo("Original", "Emphasis mode", gr.Radio, lambda: {"choices": [x.name for x in sd_emphasis.options]}, infotext="Emphasis").info("makes it possible to make model to pay (more:1.1) or (less:0.9) attention to text when you use the syntax in prompt; " + sd_emphasis.get_options_descriptions()),
"enable_batch_seeds": OptionInfo(True, "Make K-diffusion samplers produce same images in a batch as when making a single image"),
"comma_padding_backtrack": OptionInfo(20, "Prompt word wrap length limit", gr.Slider, {"minimum": 0, "maximum": 74, "step": 1}).info("in tokens - for texts shorter than specified, if they don't fit into 75 token limit, move them to the next 75 token chunk"),
"CLIP_stop_at_last_layers": OptionInfo(1, "Clip skip", gr.Slider, {"minimum": 1, "maximum": 12, "step": 1}, infotext="Clip skip").link("wiki", "https://github.com/AUTOMATIC1111/stable-diffusion-webui/wiki/Features#clip-skip").info("ignore last layers of CLIP network; 1 ignores none, 2 ignores one layer"),
|