Kolors
Stableby Kwai
Bilingual Chinese + English text-to-image model from Kwai. Uses SDXL UNet (2.6B) with ChatGLM3-6B (6.2B) as text encoder instead of CLIP, enabling strong multilingual prompt understanding. Apache 2.0 licensed.
VRAM requirements, GPU fit, and setup notes for Kolors, including 8GB/12GB fit guidance where relevant. Recommended runtimes: ComfyUI and Diffusers support. Best download size: ~6.9 GB at FP16.
- Bilingual Chinese + English prompting
- ChatGLM3-6B text encoder — not CLIP
- SDXL UNet architecture (2.6B)
- Apache 2.0 — fully open for commercial use
Your hardware
Detecting...
Image Quality Benchmarks
Measured quality metrics for Kolors outputs.
How often humans prefer this model's output (0-100%)
Visual quality and composition rating (5-9 scale)
Text-image alignment accuracy (higher is better)
VRAM Requirements by Resolution and Precision
Compare which GPUs can run Kolors at different precisions. FP8 uses less memory than FP16 when available, and the grade shows how comfortably each GPU handles the workload.
FP16 (full precision)
| Resolution | VRAM Required | RTX 4090 24GB | RTX 3060 12GB | RTX 4060 8GB | MacBook Pro M4 Pro 24GB |
|---|---|---|---|---|---|
| 512×512 | 12.7 GB | S● | B● | F● | S● |
| 768×768 | 12.8 GB | S● | B● | F● | S● |
| 1024×1024 | 13.0 GB | S● | B● | F● | S● |
Optimization Tips
Turbo / LCM distillation
Use distilled scheduler at 4-8 steps for faster iteration
ControlNets available
Add guided generation with 3 adapters (+1.5 GB VRAM each)
Run with Python
from diffusers import KolorsPipeline
import torch
pipe = KolorsPipeline.from_pretrained(
"Kwai-Kolors/Kolors",
torch_dtype=torch.float16
)
pipe.to("cuda")
image = pipe(
prompt="your prompt here",
num_inference_steps=50,
guidance_scale=5.0,
height=1024,
width=1024,
).images[0]
image.save("output.png")Get started
Setup instructions for running Kolors locally
1. Download the model
Get the checkpoint from HuggingFace
2. Place in:
ComfyUI/models/checkpoints/3. Launch ComfyUI
python main.pyComfyUI Workflow
Basic txt2img workflow for Kolors
Drag & drop into ComfyUI or use File → Import
Memory Breakdown
VRAM allocation at 1024×1024 on RTX 4090 24GB (24 GB)
Estimated Generation Time
Time per image at 1024×1024, 28 steps, FP16.
Sample Outputs
Available Formats, Downloads & Setup
Download Kolors in the precision that matches your GPU. Lower precision usually means less VRAM pressure, while higher precision keeps more quality.
| フォーマット | 精度 | サイズ | プロバイダー | |
|---|---|---|---|---|
| safetensors推奨 | FP16 | 6.9 GB | official | ダウンロード |
ControlNet Support
3 ControlNets available for Kolors. ControlNets add guided image generation (edges, depth, pose) at the cost of extra VRAM.
Canny Edge
+1.5 GB VRAMOfficial Kolors ControlNet for canny edge guidance. Works with the ChatGLM3 text encoder pipeline.
Pose
+1.5 GB VRAMOfficial human pose ControlNet for character positioning in Kolors.
LoRA Ecosystem
LimitedSmall but growing ecosystem. Kolors uses a different text encoder so SDXL LoRAs are not compatible.
Related Workflows
You might also like
Frequently asked questions
FAQ — Kolors VRAM, Runtimes & Fit
How much VRAM does Kolors need?
Kolors (2.6B parameters) requires approximately 13.0 GB of VRAM at FP16 precision for standard 1024×1024 image generation. If you want a lighter setup, lower precisions like FP8 can reduce memory use when available.
Can I run Kolors on an 8GB GPU?
Kolors usually needs more than 8GB for comfortable local use. Check the VRAM table above for the exact resolution and precision trade-off.
Does Kolors work in ComfyUI and Diffusers?
Kolors is marked for ComfyUI and Diffusers support in our catalog, so those are the runtimes we recommend first for local setup. If your workflow uses another front end, check the model's available formats and workflow notes above before downloading.
Can I run Kolors on RTX 4090?
Yes, the RTX 4090 (24 GB VRAM) can run Kolors comfortably at FP16. Expected generation time is around ~8s per image at 1024×1024.
Does Kolors support ControlNet?
Yes, Kolors has 3 ControlNet adapters available: Canny Edge, Depth Map, Pose. Each ControlNet adds roughly 1.5 GB of extra VRAM.
Does Kolors have LoRA support?
Small but growing ecosystem. Kolors uses a different text encoder so SDXL LoRAs are not compatible. The LoRA ecosystem for Kolors is rated as "minimal". Each LoRA adds roughly 0.2 GB of extra VRAM.
How fast is Kolors?
On a reference GPU (RTX 4090 24GB), Kolors generates a 1024×1024 image in approximately ~8s at FP16 with 28 inference steps. Faster GPUs with higher memory bandwidth will produce images more quickly.
About Kolors
See also