---
title: "Generate Controlled Video with ComfyUI — Troubleshooting"
canonical: "https://build.nvidia.com/playbooks/video-gen-guide/troubleshooting.md"
---

| Symptom | Cause | Fix |
|---------|-------|-----|
| Generated 3D object has a clean silhouette but messy textures | Source image had a complex or cluttered background, strong directional shadows, or a non-ideal framing | Use 1:1 aspect ratio images with a plain or removed background, neutral even lighting, and a clear front-facing or three-quarter view; pre-remove busy backgrounds before TRELLIS |
| 3D object geometry or texture quality is low | Sparse Structure / Latent Sampling steps or CFG too low for the prompt | Raise Sparse Structure Sampling Steps for cleaner topology; raise Latent Sampling Steps for surface detail; increase CFG on both; start from defaults and increase if the output does not match the prompt |
| SANA previews are slow or lower quality than expected | Preview resolution not matched to SANA’s sweet spot | Use 1024×1024 for quality; use 512 or 768 for faster iteration previews |
| Need faster image iteration than FLUX.1 | Full FLUX.1 Depth path is heavier than needed for early layout tests | In the Load Model node, try a smaller model such as SDXL; use positive/negative prompts and Wildcards in the ComfyUI graph for lighting variation |
| LTX-2.3 output looks soft or takes too long while iterating | Resolution or frame count too high for draft passes | Iterate at 1280×720 (or 1280×704) and keep sequences under 257 frames; raise to 1920×1088 when ready for final quality |
| LTX motion ignores the prompt or looks unnatural | Guidance / step settings out of a useful range | Use 20–30 steps while iterating and 40+ for finals; set Guidance Scale around 3.0–3.5 |
| Last frame does not match the provided end image | End-frame guide strength or timing too weak; long clips degrade end-frame adherence | Raise last-frame strength to 1.0; try last-frame position index `-12` instead of `-1`; keep sequences near 5 seconds (121 frames) |
| Output video is completely black | Invalid frame count, missing crop guides, or missing text encoder | Use frame counts that follow `(N×8)+1` (49, 65, 97, 121, …); for FirstFrame/LastFrame workflows add LTXVCropGuides before VAE decode; confirm the Gemma text encoder loaded |
| Subject appearance drifts mid-video | Model limitation on long or multi-motion clips | Keep clips to about 5 seconds; describe one clear motion; reduce CFG to about 3.0–3.5; for recurring characters, a subject LoRA improves consistency |
| Upscale quality or speed is off | Upscale factor or quality level not matched to source resolution | Set Upscale Factor 1–4 from input size to target (for ~720p→4K use 3); set Quality Level to 4 for maximum edge sharpening unless you need speed |
| Prompt or negative prompt fights the keyframes | Prompt restates what is already in the images, or negative list is too long | Describe change and motion, not the static look; keep negatives focused (for example: morphing, distortion, warping, flicker, jitter, blur, artifacts); LTX-2.3 does not require a negative prompt |
| GPU out-of-memory or very slow generation | Workload exceeds dedicated VRAM, or other GPU apps are running | Close other GPU applications; lower resolution, frame count, or upscale factor; monitor with Task Manager / GPU tools on Windows |

> [!NOTE]
> On discrete-GPU hardware platforms, GPU memory is separate from system RAM. Out-of-memory usually means the workload exceeds VRAM: reduce resolution, frame count, or concurrent apps. Confirm no other process is holding GPU memory before retrying.

For latest known issues, see the documentation linked under **Resources** for your hardware platform.