Sora Video Generation Limits & Glitches
Users report Sora video generation times are slower than previously. Plus plan users are temporarily limited to one concurrent generation, down from two. Generating complex scenes involving occlusion (like feet behind objects) can cause visual glitches and unnatural movement due to depth perception issues.
Fixing Model Weaknesses with Negative LoRAs
A technique surfaced for improving models like Flux: train a LoRA on undesirable outputs (e.g., plastic skin textures, AI artifacts). Applying this LoRA with negative weights during generation can force the model to avoid these traits, potentially revealing more realistic underlying capabilities.
Links:
Advanced Video Techniques and Tools
Discovered that HighRes fix techniques work well with Wan 2.1 video generation. For seamless loops, try using identical start/end frames or mirroring the generated clip. WanGP, an optimized fork, allows Wan 2.1 video generation on lower VRAM GPUs (4GB). SynCamMaster offers new camera control possibilities.
Links:
- https://jianhongbai.github.io/SynCamMaster/
- https://github.com/deepbeepmeep/Wan2GP
- https://www.reddit.com/r/StableDiffusion/comments/1j0znur/run_wan_faster_highres_fix_in_2025/
Performance Boosts and Hardware Fixes
Performance enhancements for Flux model execution include Nunchaku quantization and optimization techniques like Triton/Torch.Compile/SageAttention, installable via SwarmUI guides. RTX 5000 series GPU users can enable Stable Diffusion compatibility by installing specific nightly PyTorch builds with cu128 support.
Links:
- https://github.com/mcmonkeyprojects/SwarmUI/blob/master/docs/Advanced%20Usage.md#triton-torchcompile-sageattention-on-windows
- https://download.pytorch.org/whl/nightly/cu128
ControlNet Application Issues and Fixes
In ComfyUI, ControlNet's end_at parameter might interact unexpectedly with denoise values, causing it to stop applying earlier than set. A potential workaround involves splitting steps across two KSamplers. For ForgeUI OpenPose, set Preprocessor to "None" if using an already processed pose image.
