
1. SeedVR2 Video HD Upscale: Default configuration Best quality of 3B model (If you want to further improve the quality, you can manually select 7B model)
2. For frame interpolation: Open the corresponding group and fill in "interpolation multiplier" (Default auto-calculation: frame count * multiplier, frame rate * multiplier, ensuring audio and video synchronization with no change in original duration)
Model list is as follows:
3B Model: Faster speed, lower memory requirements
seedvr2_ema_3b_fp16.safetensorsFP16 (Best quality)
seedvr2_ema_3b_fp8_e4m3fn.safetensorsFP8 8-bit (High quality)
seedvr2_ema_3b Q4_K_M.ggufGGUF 4-bit quantization (Acceptable quality)
seedvr2_ema_3b Q8_0.ggufGGUF 8-bit quantization (High quality)
7B Model: Higher image quality, higher memory requirements
seedvr2_ema_7b_fp16.safetensorsFP16 (Best quality)
seedvr2_ema_7b_fp8_e4m3fn_mixed_block35_fp16.safetensors: FP8, last data block uses FP16 to reduce artifacts (Good quality)
seedvr2_ema_7b Q4_K_M.ggufGGUF 4-bit quantization (Acceptable quality)
seedvr2_ema_7b_sharp_*: Sharpened version, enhances details
Parameter description:
Annotations of main nodes and model parameters are mainly derived from the model and nodes [official translation], occasionally with some explanations coming from AI analysis model analogies in plain terms. If there are discrepancies, corrections are welcome, still learning...
Video HD zooming and frame interpolation: SeedVR2
1. SeedVR2 Video HD Upscale: Default configuration Best quality of 3B model (If you want to further improve the quality, you can manually select 7B model)
2. For frame interpolation: Open the corresponding group and fill in "interpolation multiplier" (Default auto-calculation: frame count * multiplier, frame rate * multiplier, ensuring audio and video synchronization with no change in original duration)
Model list is as follows:
3B Model: Faster speed, lower memory requirements
seedvr2_ema_3b_fp16.safetensorsFP16 (Best quality)
seedvr2_ema_3b_fp8_e4m3fn.safetensorsFP8 8-bit (High quality)
seedvr2_ema_3b Q4_K_M.ggufGGUF 4-bit quantization (Acceptable quality)
seedvr2_ema_3b Q8_0.ggufGGUF 8-bit quantization (High quality)
7B Model: Higher image quality, higher memory requirements
seedvr2_ema_7b_fp16.safetensorsFP16 (Best quality)
seedvr2_ema_7b_fp8_e4m3fn_mixed_block35_fp16.safetensors: FP8, last data block uses FP16 to reduce artifacts (Good quality)
seedvr2_ema_7b Q4_K_M.ggufGGUF 4-bit quantization (Acceptable quality)
seedvr2_ema_7b_sharp_*: Sharpened version, enhances details
Parameter description:
Annotations of main nodes and model parameters are mainly derived from the model and nodes [official translation], occasionally with some explanations coming from AI analysis model analogies in plain terms. If there are discrepancies, corrections are welcome, still learning...
1. SeedVR2 Video HD Upscale: Default configuration Best quality of 3B model (If you want to further improve the quality, you can manually select 7B model)
2. For frame interpolation: Open the corresponding group and fill in "interpolation multiplier" (Default auto-calculation: frame count * multiplier, frame rate * multiplier, ensuring audio and video synchronization with no change in original duration)
Model list is as follows:
3B Model: Faster speed, lower memory requirements
seedvr2_ema_3b_fp16.safetensorsFP16 (Best quality)
seedvr2_ema_3b_fp8_e4m3fn.safetensorsFP8 8-bit (High quality)
seedvr2_ema_3b Q4_K_M.ggufGGUF 4-bit quantization (Acceptable quality)
seedvr2_ema_3b Q8_0.ggufGGUF 8-bit quantization (High quality)
7B Model: Higher image quality, higher memory requirements
seedvr2_ema_7b_fp16.safetensorsFP16 (Best quality)
seedvr2_ema_7b_fp8_e4m3fn_mixed_block35_fp16.safetensors: FP8, last data block uses FP16 to reduce artifacts (Good quality)
seedvr2_ema_7b Q4_K_M.ggufGGUF 4-bit quantization (Acceptable quality)
seedvr2_ema_7b_sharp_*: Sharpened version, enhances details
Parameter description:
Annotations of main nodes and model parameters are mainly derived from the model and nodes [official translation], occasionally with some explanations coming from AI analysis model analogies in plain terms. If there are discrepancies, corrections are welcome, still learning...
