ComfyUI model file sizes: the text encoder is half the download

A character feed keeps two or three models on disk at once, and the number on a model card is the wrong one to budget with. On 19 October 2026 we added up the files ComfyUI actually loads for five current open models, read from Comfy-Org's Hugging Face repositories, and found that in two of them the text encoder is bigger than the image model itself: Qwen-Image 2.1 ships a 17.53 GB encoder next to a 14.23 GB image model, and MOSS-Image a 19.31 GB encoder next to a 10.16 GB one.

This is a measured data study, not a review. Every number below is a file size read from a public repository on one day, the arithmetic is shown, and the limits are listed at the end. Quality, speed and VRAM are a different question, and we say where we did not measure them.

What we measured and how

For each model we took the files Comfy's own repository tells you to put in models/: the diffusion model, the text encoder and the VAE. We summed them in decimal gigabytes (10^9 bytes) as the Hugging Face file listing reports them, and built three stacks per model. The bf16 stack is the full-precision diffusion file plus the full-precision encoder. The int8 stack swaps in the int8_convrot builds where Comfy lists them. The lightest stack takes the smallest diffusion file and the smallest encoder the repository lists.

Two choices are ours and you should know them. Z-Image Turbo has no int8 encoder in its repository, so its int8 stack uses the 5.63 GB fp8 encoder. MiniMax H3 is a video model with two VAEs, so its VAE column is the video VAE plus the audio VAE. LoRAs, control models and prompt-enhancer variants are left out of every stack.

huggingface.coOpen the page

The text encoder is the hidden half

Here are the bf16 stacks, file by file. The image model is the part people name, but it is not always the biggest part of the download.

The encoder share of the full-precision download is 54.0% for Qwen-Image 2.1, 64.8% for MOSS-Image, 38.9% for Z-Image Turbo, 41.7% for MiniMax H3 and 25.1% for Krea 2. Krea 2 is the one stack where the image model dominates, at 26.28 GB of a 35.41 GB total.

The filenames also show why you cannot save space by sharing. The five stacks load five different encoders: qwen3vl_8b, qwen3vl_4b, qwen3.5_9b, qwen_3_4b and qwen3vl_32b_minimax_h3. Krea 2 and Z-Image Turbo both use a roughly 4B encoder, but the files are different sizes (8.88 GB and 8.04 GB), so they are different files. Every model you add brings its own encoder.

Why the smallest image model is not the smallest download

MOSS-Image has the smallest image transformer in the set, 10.16 GB against Krea 2's 26.28 GB, which is 2.6 times smaller. Its full stack is only 1.19 times smaller (29.81 GB against 35.41 GB), because the 19.31 GB encoder takes back almost all of the difference. If you pick a model by its parameter count or the size of its main file, you will misjudge your disk.

The same pattern is why a model that looks light can be slow to start: the encoder has to be read from disk and moved before the first image. Our earlier piece on ComfyUI out of memory errors covers what happens when the loader and the card disagree about what fits.

What 8-bit and 4-bit builds do to the bill

Comfy lists int8_convrot builds for every model in the study. They roughly halve each file, so every stack lands between 52% and 59% of its bf16 size. The lightest listed combination goes further, down to 8.33 GB for Z-Image Turbo and 12.78 GB for Krea 2.

Format also changes speed, not only size. A tester running Krea 2 Turbo on one RTX 3090 measured 10.2 seconds per image with the int8 build and 18.2 seconds with fp8, 1.78 times faster, at 8 steps. That is one rig and one author, so treat it as a lead to check, not a rule. It is in the test's public README, which also reports checkpoint sizes of 12.6 GB for int8 and 12.2 GB for fp8.

The thread agrees with the table on one thing: the answer to "what do I need" depends on which build you load, and people on small cards got there by choosing a smaller file and letting ComfyUI use system RAM. Note that the thread is from June and about one model.

What a whole library costs

Disk is the budget people forget until a drive fills. Add up the four image models and then the video model and the choice of precision matters more than the choice of model.

Our verdict for a character feed

Our recommendation is to store int8 by default and keep full precision for the one model you finish with. Pick one image model and one video model as the working pair, because the study shows each extra model brings its own encoder and none of it is shared. The Comfy docs for MOSS-Image make the same split: the templates load the int8 files, and the bf16 builds are there "for more VRAM".

On disk alone Z-Image Turbo is the cheapest stack in the set (20.69 GB at bf16, 12.17 GB at int8). That says nothing about whether it keeps a face, and for that you need a test, not a file listing. Our comparison of Krea 2 and Qwen-Image 2.1 tested identity, and the explainer on distilled and base checkpoints says when a fast build is enough for drafts. For video, the sizes above explain why the MiniMax H3 review and the LTX-2.5 image-to-video guide both start with what fits.

If you run your own ComfyUI, the studio is built for this: Musetta installs what a workflow is missing and keeps each character's reference photos, so switching to a lighter stack does not change who your character is. The study holds without it, and our comparison with working in ComfyUI directly says where the manual route is enough. For the files a workflow cannot find, the missing nodes guide is the place to start.

Musetta studioDemo data

Limits of this study

• File size is not VRAM. ComfyUI can offload and unload parts of a stack, so a 32 GB download does not need 32 GB of VRAM; see the out-of-memory article above.
• One day, one source. Sizes were read from the Hugging Face file listing on 19 October 2026 and repositories change; MOSS-Image's bf16 encoder was uploaded about 20 hours before our reading.
• Comfy-Org's repackaged files only. Community GGUF and other quantizations, which some guides recommend for small cards, are not in the study.
• Quality and speed were not measured by us. The one speed figure is a third party's single-GPU test.
• Licenses are as named on the repositories. Qwen-Image 2.1 reads "qwen-research" on the Comfy-Org page and the Krea 2 card names the Krea 2 Community License without stating terms on the page we read. Read the full text before you monetize.