77 token limit
imba-pericia opened this issue · comments
The outputs you shared look amazing! Its absolutely possible to use compel just as you would use it with diffusers, here is an example
For SDXL based SegMoEs:
from compel import Compel, ReturnedEmbeddingsType
from segmoe import SegMoEPipeline
t2i = SegMoEPipeline("segmind/SegMoE-4x2-v0", device = "cuda")
compel = Compel(tokenizer=[t2i.pipe.tokenizer, t2i.pipe.tokenizer_2] , text_encoder=[t2i.pipe.text_encoder, t2i.pipe.text_encoder_2], returned_embeddings_type=ReturnedEmbeddingsType.PENULTIMATE_HIDDEN_STATES_NON_NORMALIZED, requires_pooled=[False, True])
prompt = "Milky Way. Night sky with stars and silhouette of a standing happy man with yellow light. Space background, (sharp focus:1.2), extremely detailed, (photorealistic:1.4), (RAW image, 8k high resolution:1.2), RAW candid cinema, 16mm, color graded Portra 400 film, ultra realistic, subsurface scattering, ray tracing, (volumetric lighting), extreme contrast, intricate details, reflections on ice, reflections on water, water pouring down"
conditioning, pooled = compel(prompt)
img = t2i(
prompt_embeds=conditioning,
pooled_prompt_embeds=pooled,
height=1024,
width=1024,
num_inference_steps=25,
guidance_scale=7.5,
).images[0]
img.save(f"image.png")
For SD Based SegMoEs
from segmoe import SegMoEPipeline
from compel import Compel
t2i = SegMoEPipeline("segmind/SegMoE-SD-4x2-v0", device = "cuda")
compel = Compel(tokenizer=t2i.pipe.tokenizer, text_encoder=t2i.pipe.text_encoder)
prompt = "Milky Way. Night sky with stars and silhouette of a standing happy man with yellow light. Space background, (sharp focus:1.2), extremely detailed, (photorealistic:1.4), (RAW image, 8k high resolution:1.2), RAW candid cinema, 16mm, color graded Portra 400 film, ultra realistic, subsurface scattering, ray tracing, (volumetric lighting), extreme contrast, intricate details, reflections on ice, reflections on water, water pouring down"
prompt_embeds = compel(prompt)
img = t2i(
prompt_embeds=prompt_embeds,
height=1024,
width=1024,
num_inference_steps=25,
guidance_scale=7.5,
).images[0]
img.save(f"image.png")
I hope this helps!
I hope this helps!
Thank you, it started, I will test it.
Tested it very quickly, it seemed to me that the quality had decreased, the accuracy of long prompts had increased, I took very long prompts for testing, perhaps they were initially “crooked”.
Tried to add - truncate_long_prompts = False
self.compel = Compel(
tokenizer=[self.pipeline.pipe.tokenizer, self.pipeline.pipe.tokenizer_2],
text_encoder=[self.pipeline.pipe.text_encoder, self.pipeline.pipe.text_encoder_2],
returned_embeddings_type=ReturnedEmbeddingsType.PENULTIMATE_HIDDEN_STATES_NON_NORMALIZED,
requires_pooled=[False, True],
truncate_long_prompts = False
)
And - max_embeddings_multiples=3
img = self.pipeline(
prompt_embeds=conditioning,
pooled_prompt_embeds=pooled,
height=int(height),
width=int(width),
num_inference_steps=int(num_inference_steps),
guidance_scale=float(guidance_scale),
max_embeddings_multiples=3
).images[0]
(found footage DOF, Aperture, character, hypermaximalist, slutty, beautiful, exotic, rev
ealing, appealing,:1.3), (provocative legwear:1.3), (full body photo:1.3), dynamic scene, action packed, solo, headdress, (v
oluminous petticoat:1.2 skirt black shiny satin), grand interior, baroque elements, elegant, detailed, 8k resolution, (Royal
ty:1.3), she is feeling furious, Fighter, subtle cheeks and Pouty lips and Symmetrical shaped face, in Mysterious Chaotic Tr
ansparent Unique Neon Lighting, The Chaotic Transparent Unique Neon Lighting is inspired by fantasy, pointe pose, Gray hair
styled as Bald, Cluttered Colorful Ruff, Funny Glasses, Sun in the sky, horizon-centered, Vivid (best quality, masterpiece:1
.2), photorealistic
It might be an effect of compel, having more tokens might be having a negative impact on the quality.