Full body, 16:9, 9:16, realtime video model

Here is a technical breakdown of why this happens with Hedra Realtime Avatars (Beta) and how to resolve it:
1. Architectural Differences: Batch Diffusion vs. Real-Time Streaming

• Standard Video Pipeline (Character-1 / Character-2): Offline generation relies on full frame-by-frame latent diffusion inference. Because latency is not a factor, the pipeline can run full UNet passes at high spatial resolutions (1080p+) and encode the final video using high-bitrate containers (H.264/HEVC with low CRF values).

• Realtime Avatars (Beta): To achieve glass-to-glass latencies under 500ms for live interaction, the realtime pipeline utilizes a compressed/distilled neural rendering model running at a lower native tensor resolution (typically 512x512). The output is streamed live over WebRTC.

2. High-Frequency Spatial Detail & Codec Artifacts

• Quantization & Bitrate Limits: Your source image contains an extreme density of high-frequency spatial detail (knitted wool patterns, individual hair strands, micro-snowflakes).

• WebRTC Adaptive Encoding: Real-time WebRTC encoders (H.264/VP8) dynamically adjust Quantization Parameters (QP) to preserve motion fluidity and prevent packet drop/buffering. High-frequency details consume massive spatial bandwidth; when constrained by real-time bitrate caps, the codec aggressively drops high-frequency DCT coefficients, resulting in heavy macroblocking, temporal flickering, and pixelation.

• Resampling Aliasing: When Hedra resamples your ultra-high-res source image down to the realtime model's internal input tensor, fine recurring textures (like knitwear) introduce spatial aliasing before neural inference even begins.

How to Fix / Optimize

1. Optimize the Source Image for WebRTC:
   Reduce High-Frequency Noise: Avoid source images with heavy textures (wool knitwear, snow/grain, fine mesh). Smooth surfaces, clean studio lighting, and soft backgrounds stream significantly cleaner over real-time codecs.
   Pre-Filter High-Detail Areas: If you want to use this specific character, run a subtle Gaussian blur (0.5–1px) or low-pass spatial filter on the clothing and background textures in Photoshop to soften high-frequency details. This frees up bandwidth for facial features (eyes/mouth).
   Pre-Crop to 1:1: Crop the image to a 1:1 square ratio at 1024x1024 before uploading so you control the downsampling algorithm rather than relying on automated browser/server cropping.

2. Choose the Right Pipeline for the Output:
   Use Realtime Avatars exclusively for interactive WebRTC live sessions where low latency takes priority over image fidelity.
   For non-interactive, high-definition promotional videos or showcases, use the standard Video / Character Generation pipeline in batch mode.