Rethinking Global Text Conditioning in

Diffusion Transformers

Nikita Starodubcev^1✉, Daniil Pakhomov², Zongze Wu², Ilya Drobyshevskiy¹,

Yuchen Liu², Zhonghao Wang², Yuqian Zhou²,

Zhe Lin², Dmitry Baranchuk¹

Yandex Research¹ Adobe Research^2
✉ Corresponding Author

TL;DR

⚡ CLIP pooled embeddings alone contribute little to generation quality in text-to-image and text-to-video diffusion models
We find that in state-of-the-art diffusion transformers text conditioning via modulation mechanism using pooled clip embeddings, CLIP($p$), provides little performance improvements.

⚡ Using CLIP embeddings as guidance leads to significant gains
Instead of relying on basic CLIP($p$), we propose a guidance formulation: CLIP($p$) $+ w (\text{CLIP}(p^+) - \text{CLIP}(p^-))$, where $p^+$ and $p^-$ denote positive and negative prompts, respectively.

⚡ Simple, training-free, and broadly applicable
Our method requires no additional training, incurs negligible runtime overhead, and can be applied to various tasks. It only involves selecting suitable $(p^+, p^-)$ and the guidance weight $w$.

Gallery

More image examples. Click to expand

More video examples. Click to expand

Original Model	Ours
Prompt: "A dynamic interaction between the ocean and a large rock. The rock, with its rough texture and jagged edges, is partially submerged in the water, suggesting it is a natural feature of the coastline. The water around the rock is in motion, with white foam and waves crashing against the rock, indicating the force of the ocean's movement. The background is a vast expanse of the ocean, with small ripples and waves, suggesting a moderate sea state. The overall style of the scene is a realistic depiction of a natural landscape, with a focus on the interplay between the rock and the water."

Original Model	Ours
Prompt: "A cat walks on the grass, realistic"

Original Model	Ours
Prompt: "A stylish woman walks down a Tokyo street filled with warm glowing neon and animated city signage. She wears a black leather jacket, a long red dress, and black boots, and carries a black purse. She wears sunglasses and red lipstick. She walks confidently and casually. The street is damp and reflective, creating a mirror effect of the colorful lights. Many pedestrians walk about."

Available models

• Text-to-image generation
  • FLUX schnell, detailed showcases
  • FLUX
  • HiDream
  • SD3.5 Large
  • SD3.5 Large, DMD2
  • COSMOS
• Text-to-video generation
  • HunyanVideo

FLUX schnell

Here, we discuss different showcases of the proposed approach by considering various prompts and the guidance weight. In the paper, we focus on the following directions:

1. Direction 1: Complexity
   prompt_positive = Extremely complex, the highest quality
prompt_negative = Very simple, no details at all

2. Direction 2: Aeshtetics, realism
   prompt_positive = Ultra-detailed, photorealistic, cinematic
prompt_negative = Low-res, flat, cartoonish

3. Direction 3: Hands correction
   prompt_positive = Natural and realistic hands
prompt_negative = Unnatural hands

4. Direction 4: Object counting
   prompt_positive = [n] [objects]
prompt_negative = Very simple, no details at all

5. Load the needed dependencies for the FLUX schnell model

importtypesimporttorchfromfunctoolsimportpartialfromdiffusersimportFluxPipelinefrommodels.flux_schnellimportencode_prompt, forward_modulation_guidance# Import a modelpipe=FluxPipeline.from_pretrained("black-forest-labs/FLUX.1-schnell", torch_dtype=torch.bfloat16).to('cuda')

2. Define the desired prompt, positive direction and guidance weight. We first consider the complexity direction with w = 3, start_layer = 5.

# Define the hyperparametrs: # 1. Prompts: Generation prompt, positive and negative prompts# 2. Modulation guidance strength (w)# 3. Guidance start layerprompt='A wolf on a plain background'prompt_positive='Extremely complex, the highest quality'prompt_negative='Very simple, no details at all'w=3start_layer=5# Get pooled CLIP embeddingsclip_positive=encode_prompt(pipe=pipe, prompt=prompt_positive)
clip_negative=encode_prompt(pipe=pipe, prompt=prompt_negative)
# Change forward of the pipe using the prompts and guidance weightforward_modulation_guidance=partial(forward_modulation_guidance, pooled_projections_1=clip_positive, pooled_projections_0=clip_negative,
w=w, start_layer=start_layer)
pipe.transformer.forward=types.MethodType(forward_modulation_guidance, pipe.transformer)

3. Run generation

# Run generationseed=0image=pipe([prompt] *1,
guidance_scale=0.0,
num_inference_steps=4,
max_sequence_length=256,
generator=torch.Generator("cpu").manual_seed(seed),
output_type='pil').images

Here, we provide different results for this prompt to better understand the hyperparameters

Complexity modulation guidance enhances the level of detail in image content (e.g., making a wolf’s fur more intricate), with higher values of w leading to greater complexity. A dynamic guidance strategy (start_layer = 5) increases visual richness while preserving prompt alignment. In contrast, constant guidance scales (start_layer = 0) can sometimes weaken adherence to the prompt.

We provide examples for other directions:

Note:
Please note that the hand correction and object counting directions are highly sensitive to hyperparameters.
We recommend trying different values of $w$, start_layer, and prompt
(e.g., for hand correction, adding an object often helps, such as: A boy with natural and realistic hands).

Other text-to-image models

FLUX Code. Click to expand

importtypesimporttorchfromfunctoolsimportpartialfromdiffusersimportFluxPipelinefrommodels.flux_schnellimportencode_prompt, forward_modulation_guidance# Import a modelpipe=FluxPipeline.from_pretrained("black-forest-labs/FLUX.1-dev", torch_dtype=torch.bfloat16).to('cuda')
# Define the hyperparametrs: # 1. Prompts: Generation prompt and positive, negative prompts# 2. Modulation guidance strength (w)# 3. Guidance start layerprompt="a cardboard spaceship"prompt_positive="Extremely complex, the highest quality"prompt_negative="very simple, no details at all"w=3start_layer=5# Get pooled CLIP embeddingsclip_positive=encode_prompt(pipe=pipe, prompt=prompt_positive)
clip_negative=encode_prompt(pipe=pipe, prompt=prompt_negative)
# Change forward of the pipe using the prompts and guidance weightforward_modulation_guidance=partial(forward_modulation_guidance, pooled_projections_1=clip_positive, pooled_projections_0=clip_negative,
w=w, start_layer=start_layer)
pipe.transformer.forward=types.MethodType(forward_modulation_guidance, pipe.transformer)
seed=0image=pipe([prompt] *1,
guidance_scale=3.5,
num_inference_steps=50,
max_sequence_length=512,
generator=torch.Generator("cpu").manual_seed(seed),
output_type='pil').images

HiDream Code. Click to expand

importtorchimporttypesfromfunctoolsimportpartialfromtransformersimportPreTrainedTokenizerFast, LlamaForCausalLMfromdiffusersimportHiDreamImagePipelinefrommodels.hidreamimportencode_prompt, forward_modulation_guidancetokenizer_4=PreTrainedTokenizerFast.from_pretrained("meta-llama/Meta-Llama-3.1-8B-Instruct")
text_encoder_4=LlamaForCausalLM.from_pretrained(
"meta-llama/Meta-Llama-3.1-8B-Instruct",
output_hidden_states=True,
output_attentions=True,
torch_dtype=torch.bfloat16,
)
pipe=HiDreamImagePipeline.from_pretrained(
"HiDream-ai/HiDream-I1-Fast",
tokenizer_4=tokenizer_4,
text_encoder_4=text_encoder_4,
torch_dtype=torch.bfloat16,
)
pipe=pipe.to('cuda')
# Define the hyperparametrs: # 1. Prompts: Generation prompt and positive, negative prompts# 2. Modulation guidance strength (w)# 3. Guidance start layerprompt="a cardboard spaceship"prompt_positive="Extremely complex, the highest quality"prompt_negative="very simple, no details at all"w=3start_layer=5# Get pooled CLIP embeddingsclip_positive=encode_prompt(pipe=pipe, prompt=prompt_positive)
clip_negative=encode_prompt(pipe=pipe, prompt=prompt_negative)
# Change forward of the pipe using the prompts and guidance weightforward_modulation_guidance=partial(forward_modulation_guidance, pooled_projections_1=clip_positive, pooled_projections_0=clip_negative,
w=w, start_layer=start_layer)
pipe.transformer.forward=types.MethodType(forward_modulation_guidance, pipe.transformer)
seed=0image=pipe([prompt],
height=1024,
width=1024,
guidance_scale=0.0,
num_inference_steps=16,
generator=torch.Generator("cpu").manual_seed(seed),
output_type='pil').images

SD3.5-Large Code. Click to expand

importtorchimporttypesfromdiffusersimportStableDiffusion3Pipelinefromfunctoolsimportpartialfrommodels.sd35importencode_prompt, forward_modulation_guidancepipe=StableDiffusion3Pipeline.from_pretrained("stabilityai/stable-diffusion-3.5-large", torch_dtype=torch.bfloat16)
pipe=pipe.to("cuda")
# Define the hyperparametrs: # 1. Prompts: Generation prompt and positive, negative prompts# 2. Modulation guidance strength (w)# 3. Guidance start layerprompt="a cardboard spaceship"prompt_positive="Extremely complex, the highest quality"prompt_negative="very simple, no details at all"w=2start_layer=2# Get pooled CLIP embeddingsclip_positive=encode_prompt(pipe=pipe, prompt=prompt_positive)
clip_negative=encode_prompt(pipe=pipe, prompt=prompt_negative)
# Change forward of the pipe using the prompts and guidance weightforward_modulation_guidance=partial(forward_modulation_guidance, pooled_projections_1=clip_positive, pooled_projections_0=clip_negative,
w=w, start_layer=start_layer)
pipe.transformer.forward=types.MethodType(forward_modulation_guidance, pipe.transformer)
seed=0image=pipe([prompt],
guidance_scale=3.5,
num_inference_steps=28,
generator=torch.Generator("cpu").manual_seed(seed),
output_type='pil').images

SD3.5-Large DMD2. Click to expand

importtorchimporttypesfromdiffusersimportStableDiffusion3Pipelinefromfunctoolsimportpartialfrommodels.sd35importencode_prompt, forward_modulation_guidancefrompeftimportPeftModelpipe=StableDiffusion3Pipeline.from_pretrained("stabilityai/stable-diffusion-3.5-large",
torch_dtype=torch.float16,
custom_pipeline='quickjkee/swd_pipeline')
pipe=pipe.to("cuda")
lora_path='yresearch/stable-diffusion-3.5-large-dmd2'pipe.transformer=PeftModel.from_pretrained(
pipe.transformer,
lora_path,
).to("cuda")
# Define the hyperparametrs: # 1. Prompts: Generation prompt and positive, negative prompts# 2. Modulation guidance strength (w)# 3. Guidance start layerprompt= "acardboardspaceshipprompt_positive="Extremely complex, the highest quality"prompt_negative="very simple, no details at all"w=2start_layer=2# Get pooled CLIP embeddingsclip_positive=encode_prompt(pipe=pipe, prompt=prompt_positive)
clip_negative=encode_prompt(pipe=pipe, prompt=prompt_negative)
# Change forward of the pipe using the prompts and guidance weightforward_modulation_guidance=partial(forward_modulation_guidance, pooled_projections_1=clip_positive, pooled_projections_0=clip_negative,
w=w, start_layer=start_layer)
pipe.transformer.forward=types.MethodType(forward_modulation_guidance, pipe.transformer)
seed=0sigmas= [1.0000, 0.9454, 0.8959, 0.7904, 0.7371, 0.6022, 0.0000]
scales= [128, 128, 128, 128, 128, 128]
image=pipe([prompt] *1,
sigmas=torch.tensor(sigmas).to('cuda'),
timesteps=torch.tensor(sigmas[:-1]).to('cuda') *1000,
scales=scales,
guidance_scale=0.0,
height=int(scales[0] *8),
width=int(scales[0] *8),
generator=torch.Generator("cpu").manual_seed(0),
output_type='pil').images

COSMOS. Click to expand

For the COSMOS models, we provide the code in models/cosmos, with the main entry point in generate.py.

Since the original COSMOS model does not include the CLIP model, we fine-tune it to enable this functionality.

To run generation, first download the checkpoint from:
https://huggingface.co/yresearch/cosmos-pooled

Then, create a validation folder and place the checkpoint inside it.

Next, run the generation script from models/cosmos:

. run.sh

The results will appear in the validation To modify the validation prompts, please refer to the utils.py file.

Hunyan text-to-video

importtorchimporttypesfromdiffusersimportAutoModel, HunyuanVideoPipelinefromdiffusers.quantizersimportPipelineQuantizationConfigfromdiffusers.utilsimportexport_to_videofromfunctoolsimportpartialfrommodels.hunyan_videoimportforward_modulation_guidance# quantize weights to int4 with bitsandbytespipeline_quant_config=PipelineQuantizationConfig(
quant_backend="bitsandbytes_4bit",
quant_kwargs={
"load_in_4bit": True,
"bnb_4bit_quant_type": "nf4",
"bnb_4bit_compute_dtype": torch.bfloat16
},
components_to_quantize=["transformer"]
)
pipe=HunyuanVideoPipeline.from_pretrained(
"hunyuanvideo-community/HunyuanVideo",
quantization_config=pipeline_quant_config,
torch_dtype=torch.bfloat16,
).to('cuda')
# model-offloading and tilingpipe.vae.enable_tiling()
# Define the hyperparametrs: # 1. Prompts: Generation prompt and positive, negative prompts# 2. Modulation guidance strength (w)# 3. Guidance start layerprompt="The camera follows behind a white vintage SUV with a black roof rack as it speeds up a steep dirt road surrounded by pine trees on a steep mountain slope, dust kicks up from it’s tires, the sunlight shines on the SUV as it speeds along the dirt road, casting a warm glow over the scene. The dirt road curves gently into the distance, with no other cars or vehicles in sight. The trees on either side of the road are redwoods, with patches of greenery scattered throughout. The car is seen from the rear following the curve with ease, making it seem as if it is on a rugged drive through the rugged terrain. The dirt road itself is surrounded by steep hills and mountains, with a clear blue sky above with wispy clouds."prompt_positive="Ultra-detailed, photorealistic, cinematic"prompt_negative="Low-res, bad quality, blurred"w=3.5start_layer=5# Get pooled CLIP embeddingswithtorch.no_grad():
_, clip_positive, _=pipe.encode_prompt(
prompt=prompt_positive,
)
_, clip_negative, _=pipe.encode_prompt(
prompt=prompt_negative,
)
# Change forward of the pipe using the prompts and guidance weightforward_modulation_guidance=partial(forward_modulation_guidance, pooled_projections_1=clip_positive, pooled_projections_0=clip_negative,
w=w, start_layer=start_layer)
pipe.transformer.forward=types.MethodType(forward_modulation_guidance, pipe.transformer)
seed=10video=pipe(prompt=prompt, width=512, height=320,
num_frames=61, generator=torch.Generator().manual_seed(seed),
num_inference_steps=30).frames[0]
export_to_video(video, "output_video.mp4", fps=15)

Original Model	Ours
Prompt: "The camera follows behind a white vintage SUV with a black roof rack as it speeds up a steep dirt road surrounded by pine trees on a steep mountain slope, dust kicks up from it’s tires, the sunlight shines on the SUV as it speeds along the dirt road, casting a warm glow over the scene. The dirt road curves gently into the distance, with no other cars or vehicles in sight. The trees on either side of the road are redwoods, with patches of greenery scattered throughout. The car is seen from the rear following the curve with ease, making it seem as if it is on a rugged drive through the rugged terrain. The dirt road itself is surrounded by steep hills and mountains, with a clear blue sky above with wispy clouds."