How do I speed up image generation in ComfyUI?

The biggest speed improvements come from: installing SageAttention (can cut generation time by 60%+), lowering resolution, using faster samplers like Euler or DPM++ 2M, keeping steps at 20–30, and loading ComfyUI models from an SSD or NVMe drive.

ComfyUI Troubleshooting Guide: Fix Bad Images, Slow Generation & Errors (2026)

Q: Why are my ComfyUI images blurry or bad quality?

The most common causes are: too few sampling steps (under 15), incorrect CFG value (too high or too low), a mismatched VAE, using the wrong sampler for your model, or generating at the wrong resolution. Start by checking steps (aim for 20–30), CFG (6–8 for most models), and confirm the VAE matches your checkpoint.

Q: Why does ComfyUI show a reconnecting error?

The reconnecting error usually means the browser lost its WebSocket connection to the ComfyUI backend. The most common cause is GPU VRAM exhaustion — when the backend crashes mid-generation, the browser loops on reconnecting. Fix it by restarting ComfyUI, lowering resolution or batch size, and checking for broken custom nodes.

ComfyUI Troubleshooting: Quick Fixes for Common Problems

Quick reference: Most ComfyUI problems fall into five categories — bad image quality, slow generation, VRAM errors, red/missing nodes, and a frozen or reconnecting UI. Jump to the relevant section using the table of contents above.

Running into problems with ComfyUI is completely normal — the tool is powerful but has many moving parts. This guide covers the most common issues with direct, practical fixes. Whether your images look blurry, your workflow is running slowly, or ComfyUI is showing a reconnecting error, the solution is usually one of a handful of well-known causes.

Red Nodes and Missing Custom Nodes

Red nodes are one of the most common issues when loading a workflow someone else created. They mean that one or more custom nodes required by the workflow are not installed in your ComfyUI.

Red nodes appear after loading a workflow

Fix: Open ComfyUI Manager (Menu > Manage Extensions), click 'Install Missing Custom Nodes', wait for the installation to complete, then restart ComfyUI. The nodes should turn back to their normal colour after restart.

Manager button is missing or not visible

Fix: In ComfyUI v0.4+, the Manager moved from the top toolbar to Menu > Manage Extensions. If Manager itself is not installed, go to the ComfyUI/custom_nodes folder and install it from GitHub: github.com/ltdrdata/ComfyUI-Manager

Nodes installed but still showing red after restart

Fix: Hard-refresh your browser (Ctrl+Shift+R). If still red, the node may have a dependency that failed to install — check the ComfyUI console/logs for the specific error message.

💡 Tip: See the full step-by-step guide: How to Fix Red Nodes in ComfyUI

Common Causes of Bad Image Outputs

Blurry, noisy, or distorted images are almost always caused by one of six things: steps too low, CFG too high or too low, the wrong sampler, a mismatched VAE, wrong resolution, or a weak prompt. Work through these in order.

1. Check Your Prompts

Vague or conflicting prompts produce inconsistent outputs. Be specific about subject, style, lighting, camera angle, and mood. Add a strong negative prompt to push away unwanted elements like blurry, deformed, bad anatomy, low quality.

💡 Tip: Check the model's page on Civitai or Hugging Face — most models include a recommended prompt format and example prompts that work well with that specific checkpoint.

2. Choose the Right Steps Count

Too few steps produce blurry, incomplete images. Too many steps waste time and rarely improve quality beyond a threshold — and can sometimes degrade results. For most models, 20–30 steps is the sweet spot.

Image Quality vs Sampling Steps — Sweet Spot: 20–30 Steps

ComfyUI output at 2 sampling steps — blurry, incomplete, low quality result — Steps: 2 — Too low, blurry

ComfyUI output at 30 sampling steps — clear, detailed, well-formed result — Steps: 30 — Recommended ✓

ComfyUI output at 250 sampling steps — over-processed, unnatural, degraded quality — Steps: 250 — Too high, unnatural

3. Adjust CFG (Classifier-Free Guidance)

CFG controls how strictly the model follows your prompt. Too low and the image ignores your prompt; too high and the output looks oversaturated, overcontrasted, or distorted. CFG 6–8 is the recommended starting range for most models. FLUX models typically use CFG 1 or disabled.

Image Quality vs CFG Value — Best Range: 6–8 for Most Models

ComfyUI image with low CFG (1–2) — output is random and does not follow the prompt closely — Low CFG (1–2) — prompt not followed

ComfyUI image with high CFG (15–20) — oversaturated, distorted, and unnatural output — High CFG (15–20) — distorted output

⚠️ Note: FLUX models work differently — they typically run with CFG at 1.0 or with the CFG node disabled entirely. Using high CFG values with FLUX will produce distorted results.

4. Match Your Sampler to Your Model

Different models are designed to work with specific samplers. Using the wrong sampler can produce muddy, noisy, or unnatural results. Always check the model's recommended sampler on its Civitai or Hugging Face page before generating.

SD 1.5 / SDXL: Euler a, DPM++ 2M Karras, or DPM++ SDE Karras work well.
FLUX: Use the Euler sampler with the Simple or Beta scheduler.
Turbo / Lightning LoRA models: Require specific samplers like TCD or LCM — check the model card.

ComfyUI image generated with the wrong sampler — noisy, muddy, and unnatural result — Wrong sampler — muddy output

ComfyUI image generated with the correct recommended sampler — clear and well-defined result — Correct sampler — clean output ✓

5. Use a Compatible VAE

A mismatched or corrupted VAE is one of the most common causes of washed-out colours, strange skin tones, or blurry details that persist even when other settings look correct. Always match the VAE to your checkpoint model.

SD 1.5 models: use the vae-ft-mse-840000-ema-pruned VAE.
SDXL models: use the sdxl_vae.safetensors or madebyollin/sdxl-vae-fp16-fix.
FLUX models: use the ae.safetensors VAE included with FLUX.
If the model page says "baked-in VAE", you can leave the VAE node empty or set it to "none".

6. Set the Correct Resolution

Generating at resolutions the model was not trained on causes composition problems — distorted faces, extra limbs, and broken proportions. Stick to the model's native resolution as a starting point, then upscale afterward if needed.

SD 1.5: 512×512 native. Up to 768×768 works well.
SDXL: 1024×1024 native. Also supports 1152×896, 896×1152.
FLUX: 1024×1024 native. Handles various aspect ratios well.
For larger output, generate at native resolution first then use the upscale workflow.

Speeding Up Slow Image Generation

Slow generation is usually caused by high resolution, too many steps, an inefficient sampler, or missing hardware optimisations. These fixes can dramatically cut generation time without meaningfully reducing quality.

1. Install SageAttention (Biggest Speed Gain)

SageAttention replaces the standard attention mechanism in ComfyUI with a faster implementation. It can cut generation time by 60% or more with no quality loss — it is the single highest-impact optimisation you can make.

Generation Time: SageAttention OFF vs ON (same workflow)

ComfyUI generation without SageAttention — generation time 115 seconds for the same workflow — SageAttention OFF — 115 seconds

ComfyUI generation with SageAttention enabled — generation time 40 seconds, 65% faster — SageAttention ON — 40 seconds ✓

💡 Tip: Full installation walkthrough: How to Install SageAttention in ComfyUI

2. Lower Resolution During Generation

Resolution has a large impact on generation time — doubling the resolution roughly quadruples the processing time. Generate at the model's native resolution (1024×1024 for SDXL and FLUX) and use an upscale workflow afterward for the final output.

3. Keep Steps at 20–30

Running 50–80 steps is rarely necessary and doubles or triples generation time. 20–30 steps produces near-identical quality to higher counts for most models. Turbo and Lightning LoRAs can produce good results in just 4–8 steps.

4. Use Faster Samplers

Some samplers are faster than others for the same step count. Euler and DPM++ 2M are consistently among the fastest without sacrificing quality. Heun and some Karras variants are slower due to extra function evaluations per step.

5. Limit Active ControlNets

Each active ControlNet adds a significant processing overhead to every generation step. Use only the ControlNets your workflow actually requires, and disable unused ones rather than leaving them connected but idle.

6. Check Your Hardware Setup

Storage: Load ComfyUI and models from an SSD or NVMe drive. An HDD can cause 3–5× slower model loading times.
RAM: 16 GB minimum for smooth operation. Less than 16 GB causes frequent swapping and can crash large workflows.
CPU: The CPU handles node preparation. Older CPUs slow down workflow queueing, not the generation itself.
GPU driver: Keep NVIDIA drivers updated — outdated drivers can reduce CUDA performance noticeably.

Fixing Memory Problems (VRAM and RAM)

The CUDA out of memory error is one of the most common ComfyUI crashes. It means your GPU ran out of VRAM during generation. Here are the most effective fixes.

Model	Min VRAM	Recommended VRAM	Notes
SD 1.5	4 GB	6 GB	Lightest model — good for low-VRAM GPUs
SDXL	8 GB	10 GB	Use FP8 or GGUF quantized for 8 GB GPUs
FLUX (full)	12 GB	16 GB+	Use FLUX GGUF Q4/Q5 for 8–12 GB GPUs
FLUX (GGUF Q4)	6 GB	8 GB	Quantized — good quality/VRAM trade-off
WAN / Video	12 GB	24 GB+	Split video into short clips for low VRAM

CUDA Out of Memory during generation

Fix: Lower image resolution, use a smaller or quantized model (FP8 / GGUF Q4), disable unused nodes like ControlNets or upscalers, and restart ComfyUI to clear any VRAM that was not properly freed.

CUDA Out of Memory on launch (not during generation)

Fix: Add --lowvram or --medvram to your ComfyUI launch arguments. On Windows, add them to your run_nvidia_gpu.bat file. On Linux/Mac, add them to your python main.py startup command.

Generation starts but VRAM runs out partway through

Fix: Enable image tiling in your workflow, or switch to a quantized model variant. For video workflows, reduce the number of frames per batch — process 8–16 frames at a time instead of the full clip.

VRAM stays full after generation finishes

Fix: Use the 'Unload Models' button in the menu bar to free VRAM without restarting. If VRAM stays stuck, restart ComfyUI entirely.

⚠️ Note: For FLUX models on 8–12 GB VRAM GPUs, use the GGUF Q4 or Q5 quantized variants instead of the full FP16 model. Quality is nearly identical but VRAM use drops from 24 GB to under 8 GB.

Launch flags for low-VRAM setups

# Add one of these to your ComfyUI startup command:

python main.py --lowvram       # For GPUs with 4–6 GB VRAM
python main.py --medvram       # For GPUs with 6–8 GB VRAM
python main.py --cpu           # Use CPU only (slow, last resort)

Fixing the ComfyUI Reconnecting Error

The "Reconnecting…" message means your browser lost its WebSocket connection to the ComfyUI backend server. This is one of the most-reported issues in 2026. The backend has crashed or become unresponsive — usually due to a VRAM crash mid-generation or a broken custom node.

ComfyUI stuck on 'Reconnecting…'

Fix: First, check if the ComfyUI terminal/console is still running. If it has closed or shows a crash, restart ComfyUI. Then hard-refresh your browser (Ctrl+Shift+R). If this happens repeatedly, the most likely cause is VRAM exhaustion — lower resolution or use a smaller model.

Reconnecting error starts after installing a new custom node

Fix: A broken or conflicting custom node is crashing the backend silently. Disable the newly installed node in ComfyUI Manager and restart. If the error goes away, that node is the cause — check its GitHub page for known issues or update it.

Port 8188 conflict or firewall blocking connection

Fix: Another process may be using port 8188. Try launching ComfyUI with --port 8189 to use a different port. Also check that your firewall or antivirus is not blocking the connection.

Reconnecting after clearing browser cache or switching browsers

Fix: ComfyUI uses WebSockets which require a stable connection to localhost. Switching to a different browser or disabling browser extensions that interfere with localhost connections (like some VPNs or privacy extensions) can resolve this.

Check the logs first. Open the ComfyUI Logs panel (bottom toolbar) or check your terminal window. The specific error message — CUDA out of memory, ModuleNotFoundError, Traceback — will tell you exactly what caused the crash. See the interface guide for how to access the Logs panel.

Fixing a Frozen or Unresponsive ComfyUI Interface

If the ComfyUI canvas becomes laggy, unclickable, or completely frozen, the cause is almost always browser performance, a very large workflow, or an overloaded GPU — not a server crash.

Canvas is unresponsive or extremely slow to click

Fix: Very large node setups (100+ nodes) stress the browser's rendering engine. Split the workflow into smaller, more focused workflows, or use Group nodes to collapse sections you are not actively editing.

UI freezes during generation

Fix: Live preview images update in real time and cause significant browser lag. Disable previews on nodes that don't need them — right-click the node and set Preview to Off. The generation is still running; the UI is just busy rendering previews.

UI started freezing after installing a new extension

Fix: Some custom nodes add hooks that run every frame and slow down the UI. Disable recently installed extensions in ComfyUI Manager and restart to identify which one is causing the issue.

ComfyUI performs better in a different browser

Fix: Chrome generally gives the best performance for ComfyUI. If using Firefox or Edge, switching to Chrome can improve canvas rendering speed noticeably. Disable any browser extensions (VPNs, ad blockers) that might intercept localhost traffic.

Terminal shows errors but UI appears stuck

Fix: Check the ComfyUI terminal/console window for the error message. A Python traceback there is much more informative than the frozen UI. Common causes: a node crashed mid-workflow, a model failed to load, or a Python dependency is missing.

Frequently Asked Questions

The most common causes are too few sampling steps (under 15), incorrect CFG (too high or too low — aim for 6–8 for most models), a mismatched VAE, using the wrong sampler for your model, or generating at the wrong resolution. Start by checking steps (20–30), CFG (6–8), and confirm the VAE matches your checkpoint.

The reconnecting error means the browser lost its WebSocket connection to the ComfyUI backend. The most common cause is GPU VRAM exhaustion — when the backend crashes mid-generation, the browser loops on reconnecting. Fix it by restarting ComfyUI, lowering resolution or batch size, and checking for broken custom nodes in your logs.

Red nodes mean one or more custom nodes required by the workflow are missing. Open ComfyUI Manager (Menu > Manage Extensions), click 'Install Missing Custom Nodes', wait for installation to complete, then restart ComfyUI. If nodes are still red after restart, hard-refresh your browser with Ctrl+Shift+R.

Lower your image resolution, use a smaller or quantized model (FP8 or GGUF Q4 instead of FP16), disable unused nodes like ControlNets, restart ComfyUI to clear VRAM, and add --lowvram or --medvram to your ComfyUI launch arguments if the error happens on startup.

The biggest speed improvements come from installing SageAttention (can cut time by 60%+), lowering resolution and generating at native resolution then upscaling, using fast samplers like Euler or DPM++ 2M, keeping steps at 20–30, and loading models from an SSD or NVMe drive rather than an HDD.