Run LongCat-Image Locally: ComfyUI Workflow, VRAM Requirements, and GGUF Quants

Jul 16, 2026

Searching for "LongCat ComfyUI" returns mostly tutorials for LongCat-Video — a different Meituan model. If you followed one of those and wondered why nothing matched, this guide is the correction: it covers LongCat-Image (the text-to-image and editing model) exclusively, with numbers taken from the official repositories rather than folklore.

TL;DR — LongCat-Image has native ComfyUI support since March 2026. The bf16 diffusion model is a 12.5 GB single file; the official README quotes ~17 GB for text-to-image with CPU offload enabled (~18 GB for Edit). Community GGUF quants currently go as low as 2.1 GB, but file size is not the same as runtime VRAM.

First, Know What You're Downloading

"6B parameters" undersells the real footprint. The 6B figure refers to the DiT backbone only — the full pipeline on HuggingFace is about 29.3 GB:

ComponentSize
Text encoder (Qwen-family VLM, 5 shards)16.58 GB
Diffusion transformer12.54 GB
VAE0.17 GB

LongCat-Image-Edit has the identical structure and size. Both are Apache-2.0 licensed — no extra output-usage restrictions exist in either repository, so commercial use of generated images is not restricted by the license.

ComfyUI's official changelog records native LongCat-Image support on 2026-03-05; the LongCat project announced the integration on 2026-03-22. LongCat-Image, LongCat-Image-Edit, and Edit-Turbo therefore need no custom nodes in a current build. (Older guides pointing to sooxt98/comfyui_longcat_image describe the pre-native era.)

  1. Update ComfyUI to a current build with the March 2026 native implementation.
  2. Download the repackaged weights from Comfy-Org/LongCat-Imagesplit_files/diffusion_models/ contains three bf16 files (12.54 GB each):
    • longcat_image — text-to-image
    • longcat_image_edit — instruction-based editing
    • longcat_image_edit_turbo — distilled fast editing (released 2026-02-03)
  3. Place them in ComfyUI/models/diffusion_models/, grab the matching text encoder and VAE per the comfy.org model page, and load the template workflow.

Note the Comfy-Org repackage ships bf16 only — there is no official fp8. For smaller footprints, see the GGUF path below.

VRAM Requirements (Official Numbers)

The official README quick-start enables enable_model_cpu_offload() and states:

  • Text-to-image: ~17 GB required
  • Edit: ~18 GB required

Two honest caveats. First, these figures assume CPU offload — the text encoder swaps in and out of system RAM, so you also want 32 GB+ of system memory and will trade some speed. On cards with more headroom you can keep the whole pipeline resident (pipe.to("cuda")) for faster generation. Second, no official VRAM figures exist for the quantized variants — anyone quoting exact numbers for GGUF builds is estimating from file size.

Practical mapping:

Your GPURealistic route
24 GB+ (4090/5090)bf16, full or partial offload
16 GBbf16 with CPU offload, tight but per official numbers workable for t2i
8–12 GBGGUF Q8_0 / Q4_K_M
Under 8 GBGGUF Q3/Q4, or use the cloud

Path 2: GGUF Quants for Low VRAM

Community quantizations (not official) are available from vantagewithai/LongCat-Image-GGUF — which includes ComfyUI-ready files — and stduhpf/LongCat-Image-gguf, with Edit and Edit-Turbo variants also quantized:

QuantFile sizeQuality trade-off
Q2_K2.1 GBSmallest listed build; severe quality loss, experimentation only
Q3_K_S2.76 GBNoticeable degradation, fine for drafts
Q4_K_M3.66 GBGood balance for 8–12 GB cards
Q8_06.71 GBNear-bf16 quality
BF1612.54 GBReference

Text rendering — LongCat-Image's headline strength, hitting 90.7% accuracy on an 8,105-character Chinese benchmark where Seedream 4.0 scores 58.5% and Qwen-Image 56.6% (technical report) — is typically the first thing aggressive quantization hurts. If accurate in-image text is why you chose this model, stay at Q8_0 or above.

Path 3: Plain Python (diffusers)

The model runs on the official diffusers pipeline — LongCatImagePipeline and LongCatImageEditPipeline (docs):

import torch
from diffusers import LongCatImagePipeline

pipe = LongCatImagePipeline.from_pretrained(
    "meituan-longcat/LongCat-Image", torch_dtype=torch.bfloat16
)
pipe.enable_model_cpu_offload()  # ~17 GB per official README

image = pipe(
    'A neon shop sign that says "深夜食堂 Midnight Diner", rainy street, photorealistic',
    num_inference_steps=50,
).images[0]
image.save("out.png")

When Local Isn't Worth It

Local makes sense for high volume, privacy, or LoRA experiments. For everything else, run the cost math:

  • fal.ai API: $0.13/megapixel for t2i, $0.15/MP for edit — roughly 7–8 one-megapixel images per dollar, zero setup.
  • longcat-image.org: the same models in a browser UI with free signup credits — the fastest way to check whether LongCat-Image's text rendering fits your use case before downloading 29 GB of weights. Our online guide is here.

FAQ

Is this the same as LongCat-Video / LongCat-Flash? No. Meituan's LongCat family spans chat models, video, and image. This guide covers LongCat-Image (weights released 2025-12-05) only — most misinformation online comes from mixing up family members.

Can I use outputs commercially? The model is Apache-2.0 with no additional output restrictions in the official repositories.

Does Edit support multiple reference images? The pipeline takes a single reference image per edit.


Want to compare quants against the full-precision cloud model? Generate with LongCat-Image online →

Longcat Image Team

Longcat Image Team

Run LongCat-Image Locally: ComfyUI Workflow, VRAM Requirements, and GGUF Quants | Blog