.ai

ang pinakamahuhusay na open source AI model

Maghanap

ComfyUI errors and how to fix them: red nodes, missing models, CUDA out of memory

Q: Red boxes everywhere after loading a downloaded workflow

The workflow uses custom nodes you do not have. Manager > Install Missing Custom Nodes, install all, Restart, reload the browser tab.

Q: CUDA out of memory at KSampler

Resolution, batch size or model precision exceed the card. Native resolution, batch 1, fp8 or GGUF weights, --lowvram. For video, fewer frames first, then more steps.

Paksa: Mga error at solusyon

Ni Captain

Inilathala 2026-05-12

The ComfyUI error messages people hit first, what each one means and the fix that works: missing custom nodes (red boxes), "Prompt outputs failed validation", CUDA out of memory, wrong model type in a loader, mat1 and mat2 shape errors, header deserialization, torch and xformers mismatches.

Hindi pa naisasalin ang pahinang ito, kaya ipinapakita ito sa Ingles.

Pangkalahatang-ideya

ComfyUI

reports problems in two places: red nodes on the canvas before you press Run, and a red error dialog (with a long traceback in the terminal) when a node fails during the run. Red nodes are almost always missing custom nodes or missing files. Run-time errors are almost always memory, a model put in the wrong loader, or a Python package that does not match the installed torch. This guide lists the messages in the order people meet them, with the cause and the shortest fix.

Every fix below assumes the portable Windows build or a standard git install with ComfyUI Manager. If a step says "update", use the Manager's Update All button, or git pull in the ComfyUI folder followed by pip install -r requirements.txt inside the same Python environment ComfyUI runs with (python_embeded\python.exe -m pip on the portable build).

When a workflow needs more GPU than you have, the alternative to fighting the error is to run the same graph on a hosted GPU:

BitVector

keeps hundreds of ComfyUI workflows preinstalled with their models and nodes, and

PirateDiffusion

runs hosted workflows from Telegram with /wf /run:name. Both are listed on the Apps page of this site.

Sanggunian

Pangalan	Uri	Ano ito
Red node with a title in red	symptom	A custom node the workflow uses is not installed. Manager > Install Missing Custom Nodes finds it from the workflow JSON.
Prompt outputs failed validation	error	A loader points at a file name that does not exist in your models folder, or a widget value is out of range. Open the node named in the message and pick a file from the list.
torch.cuda.OutOfMemoryError / Allocation on device	error	The model, the latent and the activations do not fit in VRAM . Lower resolution or batch, use a smaller precision (fp8, GGUF) or add --lowvram.
mat1 and mat2 shapes cannot be multiplied	error	The wrong kind of model in a loader: an SDXL LoRA on an SD 1.5 checkpoint , a Flux VAE on an SDXL graph, or a text encoder of the wrong family.
Error while deserializing header: HeaderTooLarge / MetadataIncompleteBuffer	error	The .safetensors file is incomplete or corrupt. Re-download it and compare the size or hash with the source page.
ERROR: Could not detect model type	error	A LoRA, VAE or ControlNet was loaded with Load Checkpoint, or the file is not a diffusion model at all. Use the loader that matches the file type.
xformers / torch version mismatch, DLL load failed	error	A package was installed for a different torch build. Reinstall torch, torchvision and xformers together for the same CUDA version, or remove xformers.
CLIPTextEncode: expected ... got ...	error	The text encoder does not match the model (Flux needs clip_l plus t5xxl, SDXL needs its own two encoders, Qwen and Wan ship their own).
Value not in list: ckpt_name	error	The workflow was saved with another person's file names. Pick your own file in the pulldown; the Manager can also map common names.
Server disconnected / reconnecting	symptom	The Python process crashed, most often from running out of system RAM while loading a big model. Check the terminal and watch Task Manager during the load.

Walang kailangang i-install - patakbuhin ang mga AI model sa cloud

Ang BitVector Prism ang pinakamadaling paraan para magsimula: pumili ng model, mag-type ng prompt at gumawa sa isang malinis na web app, nang walang kailangang i-configure. Available din ang BitVector sa Discord at sa web (SpyGlass).

Subukan ang BitVector

Hakbang-hakbang

Read the terminal, not only the dialog. The last lines before Traceback name the node and the file; the first line of the traceback that mentions a custom_nodes folder tells you which extension failed.
Red nodes first: open Manager, click Install Missing Custom Nodes, install everything listed, then Restart. Refresh the browser tab after the restart.
Missing files next: for every loader with a red outline or a "Value not in list" message, open the pulldown and choose a file you actually have, or download the file the workflow asks for into the right folder (models/checkpoints, models/loras, models/vae, models/clip, models/unet or diffusion_models, models/controlnet).
Check the family of every file in the graph: checkpoint, LoRA, VAE, text encoders and ControlNet must all belong to the same family. The model page on this site names the family at the top; a mix produces shape errors.
Memory errors: halve the batch size, drop the resolution to the family's native size (512 for SD 1.5, 1024 for SDXL and Flux), switch the model to an fp8 or GGUF version, and start ComfyUI with --lowvram or --novram if it still fails. Tiled VAE Decode replaces the normal decode for large images.
Package errors: run the update script of your build (update_comfyui_and_python_dependencies.bat on the portable build), or pip install -r requirements.txt in the right environment. If xformers is named, uninstall it; ComfyUI uses PyTorch attention without it.
Still broken: start ComfyUI with --disable-all-custom-nodes. If the workflow runs, re-enable extensions half at a time in the Manager to find the one that conflicts.
When the hardware is the limit, run the workflow hosted instead: open the same workflow on BitVector, or on PirateDiffusion with /workflow /show:name and /wf /run:name, and keep your local setup for what fits.

Mga halimbawa

Low-VRAM start on the portable build

.\python_embeded\python.exe -s ComfyUI\main.py --windows-standalone-build --lowvram --use-pytorch-cross-attention

Reinstall torch for CUDA 12.4 in the embedded Python

.\python_embeded\python.exe -m pip uninstall -y torch torchvision torchaudio xformers

.\python_embeded\python.exe -m pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu124

Find the extension behind a crash

python main.py --disable-all-custom-nodes

# runs? then re-enable extensions in Manager, half at a time

Mga tip

Keep one folder of models and point every UI at it with extra_model_paths.yaml, so ComfyUI,
Forge
and A1111 share files and you download each checkpoint once.
Name files by family when you save them (flux-, sdxl-, sd15-) so a mismatch is visible in the pulldown.
The Manager's "Fix node (recreate)" on a red node often repairs a workflow saved by an older version of an extension.
A GGUF version of Flux, Wan or Qwen (the Unet loader from ComfyUI-GGUF) runs on 8 to 12 GB cards with almost no quality loss at Q8 and small loss at Q5.
Set the Windows page file to system-managed on an SSD with 40 GB free: loading big models spikes system RAM and a small page file crashes the server silently.
Update ComfyUI before updating custom nodes, never the other way round; the core moves first and extensions follow.
Workflows from this site's guides list the files they need under Inputs; download them before you load the JSON.

file.

Paano ayusin

Delete and re-download; compare the file size with the source page. Use a download manager for files over 5 GB.

ImportError / DLL load failed after an update

Bakit ito nangyayari

torch, torchvision or xformers were updated to mismatched builds.

Paano ayusin

Uninstall the three and install them together from the PyTorch index for your CUDA version; drop xformers if it keeps failing.

Flux or Wan loads for minutes then the server dies

Bakit ito nangyayari

System RAM, not VRAM, ran out while the weights were read.

Paano ayusin

Close other programs, enlarge the page file, load the fp8 or GGUF version, or use the hosted copy on BitVector or PirateDiffusion.

Walang kailangang i-install - patakbuhin ang mga AI model sa cloud

Ang PirateDiffusion ay Telegram lamang at ginawa para sa mga pro: libo-libong model, LoRA at workflow na pinapatakbo ng mga chat command, na may walang limitasyong paggawa sa isang fixed-price plan.

Subukan ang PirateDiffusion

Mga tanong

Which folder does each file go in?

Checkpoints in models/checkpoints, LoRAs in models/loras, VAEs in models/vae, text encoders in models/clip (or models/text_encoders), bare diffusion models such as

Flux dev

in models/unet or models/diffusion_models, ControlNets in models/controlnet,

upscalers

in models/upscale_models.

How much VRAM do I need?

SD 1.5 runs on 4 GB, SDXL on 8 GB, Flux dev fp8 on 12 GB (GGUF Q5 on 8 GB),

Qwen-Image

and

Wan 2.2

14B on 16 to 24 GB with fp8 or GGUF,

HunyuanVideo

on 24 GB. The VRAM guide on this site has the table.

Can I run a workflow without a GPU at all?

Yes, on a hosted service: BitVector runs the ComfyUI workflows of this site preinstalled on cloud GPUs, and PirateDiffusion runs hosted workflows from Telegram. Neither needs an install.

Why does the same workflow work for someone else?

Different file names, different custom node versions, or a bigger card. Compare the Inputs list of the guide with your folders before anything else.

Does --lowvram slow things down?

Yes, moderately: it moves parts of the model between system RAM and VRAM. It is still far faster than a crash. --novram is slower again and is for 4 GB cards.