ComfyUI Black Video Output: Causes and Fixes for Wan, LTX and HunyuanVideo

Quick answer

An all-black video from ComfyUI almost always means the numbers went bad before they became pixels. Somewhere in the sampler or the VAE, values turned into NaN, "not a number". When ComfyUI converts the result to 8-bit pixels, every NaN becomes 0, and 0 is black. The console usually says so with one line:

RuntimeWarning: invalid value encountered in cast

The fixes that maintainers have confirmed, cheapest first:

1. Remove --use-sage-attention from your launch command, and --fast fp16_accumulation if you added it. This is the most common cause on Wan.
2. Run the VAE in full precision with --fp32-vae. ComfyUI's own help text says --fp16-vae "might cause black images".
3. Check the VAE matches the model. Wan 2.2 5B needs wan2.2_vae; the Wan 2.2 14B models use wan_2.1_vae.
4. Use plain VAE Decode, not VAE Decode (Tiled), on Wan.
5. On a GTX 16-series card or on a Mac, read the hardware section below; ComfyUI already works around both, so something may be overriding it.

Change one thing at a time and run the same seed, so you know which change fixed it.

We read ComfyUI's source, its documentation and the issue threads cited below on 2026-10-09. ComfyUI's official troubleshooting pages do not cover black output; everything here comes from the code and from maintainers' replies in issues.

Find the stage that went black

Before changing flags, find out where the frames turn black.

• Read the console. The invalid value encountered in cast warning, usually from the save node or VideoHelperSuite, confirms NaN. No warning and still black points elsewhere, such as an empty or wrong input image.
• Save an image straight after VAE Decode. Add a Save Image node on the decoded frames. If that PNG is black too, the video encoder is not the problem. If the PNG is fine, look at the video save node and its format.
• Watch the sampler preview. If the preview goes black or flat partway through sampling, the NaN came from the model or attention. If sampling previews look normal and only the decode is black, it is the VAE.
• Rerun with --disable-all-custom-nodes on a built-in template. If it still goes black, a custom node is not the cause.

Causes and fixes

The GTX 16-series list in ComfyUI's code covers the GTX 1660, 1650 and 1630, the MX450 and MX550, and the T500, T550, T600, T1000, T1200 and T2000.

The flags, spelled exactly

From comfy/cli_args.py on ComfyUI's master branch, 2026-10-09. A wrong spelling is ignored or fails, so copy them.

Two help texts in that file mention black images directly: --fp16-vae "might cause black images", and --force-upcast-attention asks you to "report if it fixes black images".

--cpu-vae runs the decode on the processor. It is slow, but if the frames come out right on the CPU, the VAE's precision on your GPU is the problem.

By model

Wan 2.1 and Wan 2.2

Start with SageAttention; it accounts for most reports. Then check the VAE pair: the 5B and the 14B use different VAEs, and the Wan 2.2 models folder page lists which template loads which.

One user in the SageAttention thread also found that one lightx2v 4-step fp8 file gave black output while its _comfyui_ variant did not. That is a single report, but if you downloaded a lightx2v file outside the ComfyUI repack, try the repack's copy.

LTX-2 and later

LTX failures are more often noise or static than pure black:

• Pure noise: the dev model run with the distilled 8-step settings and without the distilled LoRA. Load the distilled model instead.
• Static: a tutorial reports that the GGUF dual text-encoder loader left on type sd3 instead of ltxv fills the video with static.

Both are covered with the 12GB setup on our LTX Video 12GB page.

HunyuanVideo and HunyuanVideo 1.5

The original HunyuanVideo's black output in kijai's wrapper came down to the fp16 VAE and the attention mode. HunyuanVideo 1.5's ComfyUI files ship the VAE as hunyuanvideo15_vae_fp16; if it goes black, --fp32-vae is the first thing to try.

Reports nobody has solved yet

• Wan 2.2 image-to-video on an RTX 5070 Ti. The official template gives a black video and a black PNG after VAE Decode, even with all custom nodes disabled, on PyTorch 2.13 with CUDA 13.0. The issue is open, with no reply. NaN is the reporter's guess, not a confirmed cause.
• HunyuanVideo 1.5 on Apple silicon. On an M2 Ultra, 73 frames came out fine, 81 to 97 frames came out black, and 121 frames failed with a tensor-size error. A tiled decode was suggested but not confirmed to work.

If your case matches one of these, adding your hardware and versions to the open thread helps more than a new one.

When it is not black but broken

Black output is a numbers problem. A missing file, a wrong folder or a memory error stops the run instead of finishing it black. For those:

• Our ComfyUI troubleshooting page covers out-of-memory, slow steps, freezes and missing models, written for MiniMax H3.
• Our ComfyUI download page covers installs and where each one keeps its models.

GenVidKit is an independent guide. It is not affiliated with Comfy Org, Alibaba, the Wan team, Lightricks, Tencent, Kijai or Hugging Face.

Sources

All read on 2026-10-09.

• ComfyUI comfy/cli_args.py — flag names and help text.
• ComfyUI comfy/model_management.py — the GTX 16-series and macOS workarounds.
• ComfyUI issue #10007 — Wan 2.2 black output and SageAttention.
• WanVideoWrapper issue #1339 — NaN latents, SageAttention and PyTorch version.
• ComfyUI issue #7116 — tiled VAE decode on Wan.
• HunyuanVideoWrapper issue #55 — fp16 VAE and attention.
• ComfyUI issue #884 — GTX 16-series black images.
• ComfyUI issue #14892 and discussion #15109 — the unsolved reports.
• ComfyUI-LTXVideo issue #465 — LTX noise from the dev model.
• Wan2.2 tutorial and HunyuanVideo 1.5 tutorial, ComfyUI documentation — VAE pairing and files.
• ComfyUI troubleshooting documentation — checked for black output; not covered.