feat: support cache for LongCat-Image

[e1ijah1]

#587

[DefTruth]

@e1ijah1 Hi~ Can you share the image results w/ or w/o cache?

[DefTruth]

Thanks for your contribution!

[e1ijah1]

@e1ijah1 Hi~ Can you share the image results w/ or w/o cache?

Below is the longcat_image_edit.1024x1024.C0_Q0_NONE.png generated by python3 generate.py generate longcat_image_edit

Summary:

INFO 12-22 06:37:00 [base.py:557] ----------------------------------------------------------------------------------------------------
INFO 12-22 06:37:00 [base.py:342] 🤖 Example Init Config Summary:
INFO 12-22 06:37:00 [base.py:360] - Model: /data/LongCat-Image-Edit/ + meituan-longcat/LongCat-Image-Edit
INFO 12-22 06:37:00 [base.py:360] - Task Type: IE2I - Image Editing to Image
INFO 12-22 06:37:00 [base.py:360] - Torch Dtype: torch.bfloat16
INFO 12-22 06:37:00 [base.py:360] - LoRA Weights: None
INFO 12-22 06:37:00 [base.py:196] 🤖 Example Input Summary:
INFO 12-22 06:37:00 [base.py:196] - prompt: Turn the cat into a dog
INFO 12-22 06:37:00 [base.py:196] - negative_prompt:
INFO 12-22 06:37:00 [base.py:196] - guidance_scale: 4.5
INFO 12-22 06:37:00 [base.py:196] - num_inference_steps: 50
INFO 12-22 06:37:00 [base.py:196] - image: Single Image (1024x1024)
INFO 12-22 06:37:00 [base.py:196] - generator: device cpu, seed 0
INFO 12-22 06:37:00 [base.py:259] 🤖 Example Output Summary:
INFO 12-22 06:37:00 [base.py:270] - Model: longcat_image_edit
INFO 12-22 06:37:00 [base.py:270] - Optimization: C0_Q0_NONE
INFO 12-22 06:37:00 [base.py:270] - Load Time: 0.99s
INFO 12-22 06:37:00 [base.py:270] - Warmup Time: 64.54s
INFO 12-22 06:37:00 [base.py:270] - Inference Time: 54.15s
INFO 12-22 06:37:01 [base.py:227] Image saved to longcat_image_edit.1024x1024.C0_Q0_NONE.png
INFO 12-22 06:37:01 [base.py:568] ----------------------------------------------------------------------------------------------------

Below is the longcat_image_edit.1024x1024.C0_Q0_DBCache_F1B0_W8I1M0MC3_R0.24_CFG1_T0O0_S32.png generated by python3 generate.py generate longcat_image_edit --cache

Summary:

INFO 12-22 06:43:32 [base.py:557] ----------------------------------------------------------------------------------------------------
INFO 12-22 06:43:32 [base.py:342] 🤖 Example Init Config Summary:
INFO 12-22 06:43:32 [base.py:360] - Model: /data/LongCat-Image-Edit/ + meituan-longcat/LongCat-Image-Edit
INFO 12-22 06:43:32 [base.py:360] - Task Type: IE2I - Image Editing to Image
INFO 12-22 06:43:32 [base.py:360] - Torch Dtype: torch.bfloat16
INFO 12-22 06:43:32 [base.py:360] - LoRA Weights: None
INFO 12-22 06:43:32 [base.py:196] 🤖 Example Input Summary:
INFO 12-22 06:43:32 [base.py:196] - prompt: Turn the cat into a dog
INFO 12-22 06:43:32 [base.py:196] - negative_prompt:
INFO 12-22 06:43:32 [base.py:196] - guidance_scale: 4.5
INFO 12-22 06:43:32 [base.py:196] - num_inference_steps: 50
INFO 12-22 06:43:32 [base.py:196] - image: Single Image (1024x1024)
INFO 12-22 06:43:32 [base.py:196] - generator: device cpu, seed 0
INFO 12-22 06:43:32 [base.py:259] 🤖 Example Output Summary:
INFO 12-22 06:43:32 [base.py:270] - Model: longcat_image_edit
INFO 12-22 06:43:32 [base.py:270] - Optimization: C0_Q0_DBCache_F1B0_W8I1M0MC3_R0.24_CFG1_T0O0_S32
INFO 12-22 06:43:32 [base.py:270] - Load Time: 0.99s
INFO 12-22 06:43:32 [base.py:270] - Warmup Time: 35.16s
INFO 12-22 06:43:32 [base.py:270] - Inference Time: 22.47s
INFO 12-22 06:43:32 [base.py:227] Image saved to longcat_image_edit.1024x1024.C0_Q0_DBCache_F1B0_W8I1M0MC3_R0.24_CFG1_T0O0_S32.png
INFO 12-22 06:43:32 [base.py:568] ----------------------------------------------------------------------------------------------------

BTW, I'm still downloading the LongCat-Image weights. I'll post the generated images here later.

[e1ijah1]

@e1ijah1 Hi~ Can you share the image results w/ or w/o cache?

The image generated with out cache:

Summary:

INFO 12-22 18:40:26 [base.py:557] ----------------------------------------------------------------------------------------------------
INFO 12-22 18:40:26 [base.py:342] 🤖 Example Init Config Summary:
INFO 12-22 18:40:26 [base.py:360] - Model: /data/LongCat-Image + meituan-longcat/LongCat-Image
INFO 12-22 18:40:26 [base.py:360] - Task Type: T2I - Text to Image
INFO 12-22 18:40:26 [base.py:360] - Torch Dtype: torch.bfloat16
INFO 12-22 18:40:26 [base.py:360] - LoRA Weights: None
INFO 12-22 18:40:26 [base.py:196] 🤖 Example Input Summary:
INFO 12-22 18:40:26 [base.py:196] - prompt: A young Asian woman wearing a yellow knit sweater paired with a white necklace. Her hands rest on her knees, with a serene expression. The background features a rough brick wall, with warm afternoon sunlight casting upon her, creating a tranquil and cozy atmosphere. The shot uses a medium-distance perspective, highlighting her demeanor and the details of her attire. Soft lighting illuminates her face, emphasizing her facial features and the texture of her accessories, adding depth and warmth to the image. The overall composition is simple and elegant, with the brick wall's texture complementing the interplay of sunlight and shadows, showcasing the character's grace and composure.
INFO 12-22 18:40:26 [base.py:196] - height: 1024
INFO 12-22 18:40:26 [base.py:196] - width: 1024
INFO 12-22 18:40:26 [base.py:196] - guidance_scale: 4.5
INFO 12-22 18:40:26 [base.py:196] - num_inference_steps: 50
INFO 12-22 18:40:26 [base.py:196] - generator: device cpu, seed 0
INFO 12-22 18:40:26 [base.py:259] 🤖 Example Output Summary:
INFO 12-22 18:40:26 [base.py:270] - Model: longcat_image
INFO 12-22 18:40:26 [base.py:270] - Optimization: C0_Q0_NONE
INFO 12-22 18:40:26 [base.py:270] - Load Time: 0.98s
INFO 12-22 18:40:26 [base.py:270] - Warmup Time: 37.35s
INFO 12-22 18:40:26 [base.py:270] - Inference Time: 26.83s
INFO 12-22 18:40:26 [base.py:227] Image saved to longcat_image.1024x1024.C0_Q0_NONE.png
INFO 12-22 18:40:26 [base.py:568] ----------------------------------------------------------------------------------------------------

The image generated with cache:

Summary:

INFO 12-22 18:45:37 [base.py:557] ----------------------------------------------------------------------------------------------------
INFO 12-22 18:45:37 [base.py:342] 🤖 Example Init Config Summary:
INFO 12-22 18:45:37 [base.py:360] - Model: /data/LongCat-Image + meituan-longcat/LongCat-Image
INFO 12-22 18:45:37 [base.py:360] - Task Type: T2I - Text to Image
INFO 12-22 18:45:37 [base.py:360] - Torch Dtype: torch.bfloat16
INFO 12-22 18:45:37 [base.py:360] - LoRA Weights: None
INFO 12-22 18:45:37 [base.py:196] 🤖 Example Input Summary:
INFO 12-22 18:45:37 [base.py:196] - prompt: A young Asian woman wearing a yellow knit sweater paired with a white necklace. Her hands rest on her knees, with a serene expression. The background features a rough brick wall, with warm afternoon sunlight casting upon her, creating a tranquil and cozy atmosphere. The shot uses a medium-distance perspective, highlighting her demeanor and the details of her attire. Soft lighting illuminates her face, emphasizing her facial features and the texture of her accessories, adding depth and warmth to the image. The overall composition is simple and elegant, with the brick wall's texture complementing the interplay of sunlight and shadows, showcasing the character's grace and composure.
INFO 12-22 18:45:37 [base.py:196] - height: 1024
INFO 12-22 18:45:37 [base.py:196] - width: 1024
INFO 12-22 18:45:37 [base.py:196] - guidance_scale: 4.5
INFO 12-22 18:45:37 [base.py:196] - num_inference_steps: 50
INFO 12-22 18:45:37 [base.py:196] - generator: device cpu, seed 0
INFO 12-22 18:45:37 [base.py:259] 🤖 Example Output Summary:
INFO 12-22 18:45:37 [base.py:270] - Model: longcat_image
INFO 12-22 18:45:37 [base.py:270] - Optimization: C0_Q0_DBCache_F1B0_W8I1M0MC3_R0.24_CFG1_T0O0_S31
INFO 12-22 18:45:37 [base.py:270] - Load Time: 0.98s
INFO 12-22 18:45:37 [base.py:270] - Warmup Time: 24.92s
INFO 12-22 18:45:37 [base.py:270] - Inference Time: 13.51s
INFO 12-22 18:45:37 [base.py:227] Image saved to longcat_image.1024x1024.C0_Q0_DBCache_F1B0_W8I1M0MC3_R0.24_CFG1_T0O0_S31.png
INFO 12-22 18:45:37 [base.py:568] ----------------------------------------------------------------------------------------------------

[DefTruth]

LGTM~ Thanks for your contribution!