seedvr2x's model files: SeedVR2 7B, sharp 7B and VAE in float16, TransNetV2, and both 7Bs in fp8, int8, NVFP4, GGUF Q8_0, Q4_K and dynamic, with the importance matrices

c14a2bc verified 14 days ago

Raw History Blame Contribute Delete

16.9 kB

seedvr2x model files
Copyright 2026 Edouard Hur

These files are unofficial conversions made for seedvr2x, not their authors' releases. Each
entry gives a file's origin, its copyright and licence, and what was changed.

The GGUF files were quantized by ggml's own quantizer (ggml / llama.cpp, MIT License,
Copyright (c) 2023-2026 The ggml authors), built from llama.cpp at abeada3 and called
unchanged; no ggml code is in the files.

seedvr2x_ema_7b_Q4_K.gguf
  Made from: seedvr2_ema_7b.pth of ByteDance-Seed/SeedVR2-7B, revision eb0c4281d41ba3767d4f14370f0e37e9e9180c16
    https://huggingface.co/ByteDance-Seed/SeedVR2-7B/resolve/eb0c4281d41ba3767d4f14370f0e37e9e9180c16/seedvr2_ema_7b.pth
    SHA-256: e1b2ae25505607e61f2a7dc7967ba778aaf3e3626d9969ce6e24c52d9ddebfcd
  SeedVR2: Copyright (c) 2025 Bytedance Ltd. and/or its affiliates, under the Apache License, Version 2.0 (LICENSE).
  Changed: the 288 attention and MLP matrices of the blocks quantized to Q4_K by ggml_quantize_chunk (llama.cpp abeada335e2e78bd3fe63febafab7e900ce75810, no importance matrix) from the float32 master; every other tensor rounded to the nearest float16.

seedvr2x_ema_7b_Q4_K_imatrix.gguf
  Made from: seedvr2_ema_7b.pth of ByteDance-Seed/SeedVR2-7B, revision eb0c4281d41ba3767d4f14370f0e37e9e9180c16
    https://huggingface.co/ByteDance-Seed/SeedVR2-7B/resolve/eb0c4281d41ba3767d4f14370f0e37e9e9180c16/seedvr2_ema_7b.pth
    SHA-256: e1b2ae25505607e61f2a7dc7967ba778aaf3e3626d9969ce6e24c52d9ddebfcd
  SeedVR2: Copyright (c) 2025 Bytedance Ltd. and/or its affiliates, under the Apache License, Version 2.0 (LICENSE).
  Changed: the 288 attention and MLP matrices of the blocks quantized by ggml_quantize_chunk (llama.cpp abeada335e2e78bd3fe63febafab7e900ce75810) with an importance matrix (each input channel's mean square over the calibration runs of seedvr2_ema_7b_fp16.safetensors in numz's SeedVR2) from the float32 master, to Q4_K; every other tensor rounded to the nearest float16.

seedvr2x_ema_7b_Q8_0.gguf
  Made from: seedvr2_ema_7b.pth of ByteDance-Seed/SeedVR2-7B, revision eb0c4281d41ba3767d4f14370f0e37e9e9180c16
    https://huggingface.co/ByteDance-Seed/SeedVR2-7B/resolve/eb0c4281d41ba3767d4f14370f0e37e9e9180c16/seedvr2_ema_7b.pth
    SHA-256: e1b2ae25505607e61f2a7dc7967ba778aaf3e3626d9969ce6e24c52d9ddebfcd
  SeedVR2: Copyright (c) 2025 Bytedance Ltd. and/or its affiliates, under the Apache License, Version 2.0 (LICENSE).
  Changed: the 288 attention and MLP matrices of the blocks quantized to Q8_0 by ggml_quantize_chunk (llama.cpp abeada335e2e78bd3fe63febafab7e900ce75810, no importance matrix) from the float32 master; every other tensor rounded to the nearest float16.

seedvr2x_ema_7b_dyn.gguf
  Made from: seedvr2_ema_7b.pth of ByteDance-Seed/SeedVR2-7B, revision eb0c4281d41ba3767d4f14370f0e37e9e9180c16
    https://huggingface.co/ByteDance-Seed/SeedVR2-7B/resolve/eb0c4281d41ba3767d4f14370f0e37e9e9180c16/seedvr2_ema_7b.pth
    SHA-256: e1b2ae25505607e61f2a7dc7967ba778aaf3e3626d9969ce6e24c52d9ddebfcd
  SeedVR2: Copyright (c) 2025 Bytedance Ltd. and/or its affiliates, under the Apache License, Version 2.0 (LICENSE).
  Changed: the 288 attention and MLP matrices of the blocks quantized by ggml_quantize_chunk (llama.cpp abeada335e2e78bd3fe63febafab7e900ce75810) with an importance matrix (each input channel's mean square over the calibration runs of seedvr2_ema_7b_fp16.safetensors in numz's SeedVR2) from the float32 master, each to the type among Q3_K, Q4_K, Q5_K, Q6_K, Q8_0 that minimises the summed activation-weighted squared relative error of the 288 under the bytes of the uniform Q4_K; every other tensor rounded to the nearest float16.

seedvr2x_ema_7b_fp16.safetensors
  Made from: seedvr2_ema_7b.pth of ByteDance-Seed/SeedVR2-7B, revision eb0c4281d41ba3767d4f14370f0e37e9e9180c16
    https://huggingface.co/ByteDance-Seed/SeedVR2-7B/resolve/eb0c4281d41ba3767d4f14370f0e37e9e9180c16/seedvr2_ema_7b.pth
    SHA-256: e1b2ae25505607e61f2a7dc7967ba778aaf3e3626d9969ce6e24c52d9ddebfcd
  SeedVR2: Copyright (c) 2025 Bytedance Ltd. and/or its affiliates, under the Apache License, Version 2.0 (LICENSE).
  Changed: every tensor rounded from float32 to the nearest float16, ties to even; names and shapes unchanged.

seedvr2x_ema_7b_fp8_scaled.safetensors
  Made from: seedvr2_ema_7b.pth of ByteDance-Seed/SeedVR2-7B, revision eb0c4281d41ba3767d4f14370f0e37e9e9180c16
    https://huggingface.co/ByteDance-Seed/SeedVR2-7B/resolve/eb0c4281d41ba3767d4f14370f0e37e9e9180c16/seedvr2_ema_7b.pth
    SHA-256: e1b2ae25505607e61f2a7dc7967ba778aaf3e3626d9969ce6e24c52d9ddebfcd
  SeedVR2: Copyright (c) 2025 Bytedance Ltd. and/or its affiliates, under the Apache License, Version 2.0 (LICENSE).
  Changed: the 288 attention and MLP matrices of the blocks in float8_e4m3fn with one float32 scale per tensor (max|W| / 448; comfy-kitchen's FP8 layout, <layer>.weight and <layer>.weight_scale, marked by <layer>.comfy_quant), rounded to nearest, ties to even; every other tensor rounded to the nearest float16.

seedvr2x_ema_7b_int8_convrot.safetensors
  Made from: seedvr2_ema_7b.pth of ByteDance-Seed/SeedVR2-7B, revision eb0c4281d41ba3767d4f14370f0e37e9e9180c16
    https://huggingface.co/ByteDance-Seed/SeedVR2-7B/resolve/eb0c4281d41ba3767d4f14370f0e37e9e9180c16/seedvr2_ema_7b.pth
    SHA-256: e1b2ae25505607e61f2a7dc7967ba778aaf3e3626d9969ce6e24c52d9ddebfcd
  SeedVR2: Copyright (c) 2025 Bytedance Ltd. and/or its affiliates, under the Apache License, Version 2.0 (LICENSE).
  Changed: the 288 attention and MLP matrices of the blocks in int8 with comfy-kitchen's rotation (each 256 columns times the regular Hadamard / 16) and one float32 scale per row (max|row| / 127), rounded to nearest, ties to even (comfy-kitchen's INT8 layout, <layer>.weight and <layer>.weight_scale, marked by <layer>.comfy_quant); every other tensor rounded to the nearest float16.

seedvr2x_ema_7b_nvfp4.safetensors
  Made from: seedvr2_ema_7b.pth of ByteDance-Seed/SeedVR2-7B, revision eb0c4281d41ba3767d4f14370f0e37e9e9180c16
    https://huggingface.co/ByteDance-Seed/SeedVR2-7B/resolve/eb0c4281d41ba3767d4f14370f0e37e9e9180c16/seedvr2_ema_7b.pth
    SHA-256: e1b2ae25505607e61f2a7dc7967ba778aaf3e3626d9969ce6e24c52d9ddebfcd
  SeedVR2: Copyright (c) 2025 Bytedance Ltd. and/or its affiliates, under the Apache License, Version 2.0 (LICENSE).
  Changed: the 288 attention and MLP matrices of the blocks in NVFP4: E2M1 values, an E4M3 scale per 16 values along the input and a float32 scale per tensor (max|W| / (448 x 6)), each block's scale the one, of its max's (max / 6, to the nearest E4M3) and the 7 E4M3 values below it, that minimises the block's squared error, values rounded to nearest, ties to even (comfy-kitchen's NVFP4 layout, <layer>.weight, <layer>.weight_scale and <layer>.weight_scale_2, marked by <layer>.comfy_quant); every other tensor rounded to the nearest float16.

seedvr2x_ema_7b_sharp_Q4_K.gguf
  Made from: seedvr2_ema_7b_sharp.pth of ByteDance-Seed/SeedVR2-7B, revision eb0c4281d41ba3767d4f14370f0e37e9e9180c16
    https://huggingface.co/ByteDance-Seed/SeedVR2-7B/resolve/eb0c4281d41ba3767d4f14370f0e37e9e9180c16/seedvr2_ema_7b_sharp.pth
    SHA-256: ced5706c976d5879efcab9e108349d67abcbd8a9b36a1f48bf0f19c24164a264
  SeedVR2: Copyright (c) 2025 Bytedance Ltd. and/or its affiliates, under the Apache License, Version 2.0 (LICENSE).
  Changed: the 288 attention and MLP matrices of the blocks quantized to Q4_K by ggml_quantize_chunk (llama.cpp abeada335e2e78bd3fe63febafab7e900ce75810, no importance matrix) from the float32 master; every other tensor rounded to the nearest float16.

seedvr2x_ema_7b_sharp_Q4_K_imatrix.gguf
  Made from: seedvr2_ema_7b_sharp.pth of ByteDance-Seed/SeedVR2-7B, revision eb0c4281d41ba3767d4f14370f0e37e9e9180c16
    https://huggingface.co/ByteDance-Seed/SeedVR2-7B/resolve/eb0c4281d41ba3767d4f14370f0e37e9e9180c16/seedvr2_ema_7b_sharp.pth
    SHA-256: ced5706c976d5879efcab9e108349d67abcbd8a9b36a1f48bf0f19c24164a264
  SeedVR2: Copyright (c) 2025 Bytedance Ltd. and/or its affiliates, under the Apache License, Version 2.0 (LICENSE).
  Changed: the 288 attention and MLP matrices of the blocks quantized by ggml_quantize_chunk (llama.cpp abeada335e2e78bd3fe63febafab7e900ce75810) with an importance matrix (each input channel's mean square over the calibration runs of seedvr2_ema_7b_sharp_fp16.safetensors in numz's SeedVR2) from the float32 master, to Q4_K; every other tensor rounded to the nearest float16.

seedvr2x_ema_7b_sharp_Q8_0.gguf
  Made from: seedvr2_ema_7b_sharp.pth of ByteDance-Seed/SeedVR2-7B, revision eb0c4281d41ba3767d4f14370f0e37e9e9180c16
    https://huggingface.co/ByteDance-Seed/SeedVR2-7B/resolve/eb0c4281d41ba3767d4f14370f0e37e9e9180c16/seedvr2_ema_7b_sharp.pth
    SHA-256: ced5706c976d5879efcab9e108349d67abcbd8a9b36a1f48bf0f19c24164a264
  SeedVR2: Copyright (c) 2025 Bytedance Ltd. and/or its affiliates, under the Apache License, Version 2.0 (LICENSE).
  Changed: the 288 attention and MLP matrices of the blocks quantized to Q8_0 by ggml_quantize_chunk (llama.cpp abeada335e2e78bd3fe63febafab7e900ce75810, no importance matrix) from the float32 master; every other tensor rounded to the nearest float16.

seedvr2x_ema_7b_sharp_dyn.gguf
  Made from: seedvr2_ema_7b_sharp.pth of ByteDance-Seed/SeedVR2-7B, revision eb0c4281d41ba3767d4f14370f0e37e9e9180c16
    https://huggingface.co/ByteDance-Seed/SeedVR2-7B/resolve/eb0c4281d41ba3767d4f14370f0e37e9e9180c16/seedvr2_ema_7b_sharp.pth
    SHA-256: ced5706c976d5879efcab9e108349d67abcbd8a9b36a1f48bf0f19c24164a264
  SeedVR2: Copyright (c) 2025 Bytedance Ltd. and/or its affiliates, under the Apache License, Version 2.0 (LICENSE).
  Changed: the 288 attention and MLP matrices of the blocks quantized by ggml_quantize_chunk (llama.cpp abeada335e2e78bd3fe63febafab7e900ce75810) with an importance matrix (each input channel's mean square over the calibration runs of seedvr2_ema_7b_sharp_fp16.safetensors in numz's SeedVR2) from the float32 master, each to the type among Q3_K, Q4_K, Q5_K, Q6_K, Q8_0 that minimises the summed activation-weighted squared relative error of the 288 under the bytes of the uniform Q4_K; every other tensor rounded to the nearest float16.

seedvr2x_ema_7b_sharp_fp16.safetensors
  Made from: seedvr2_ema_7b_sharp.pth of ByteDance-Seed/SeedVR2-7B, revision eb0c4281d41ba3767d4f14370f0e37e9e9180c16
    https://huggingface.co/ByteDance-Seed/SeedVR2-7B/resolve/eb0c4281d41ba3767d4f14370f0e37e9e9180c16/seedvr2_ema_7b_sharp.pth
    SHA-256: ced5706c976d5879efcab9e108349d67abcbd8a9b36a1f48bf0f19c24164a264
  SeedVR2: Copyright (c) 2025 Bytedance Ltd. and/or its affiliates, under the Apache License, Version 2.0 (LICENSE).
  Changed: every tensor rounded from float32 to the nearest float16, ties to even; names and shapes unchanged.

seedvr2x_ema_7b_sharp_fp8_scaled.safetensors
  Made from: seedvr2_ema_7b_sharp.pth of ByteDance-Seed/SeedVR2-7B, revision eb0c4281d41ba3767d4f14370f0e37e9e9180c16
    https://huggingface.co/ByteDance-Seed/SeedVR2-7B/resolve/eb0c4281d41ba3767d4f14370f0e37e9e9180c16/seedvr2_ema_7b_sharp.pth
    SHA-256: ced5706c976d5879efcab9e108349d67abcbd8a9b36a1f48bf0f19c24164a264
  SeedVR2: Copyright (c) 2025 Bytedance Ltd. and/or its affiliates, under the Apache License, Version 2.0 (LICENSE).
  Changed: the 288 attention and MLP matrices of the blocks in float8_e4m3fn with one float32 scale per tensor (max|W| / 448; comfy-kitchen's FP8 layout, <layer>.weight and <layer>.weight_scale, marked by <layer>.comfy_quant), rounded to nearest, ties to even; every other tensor rounded to the nearest float16.

seedvr2x_ema_7b_sharp_int8_convrot.safetensors
  Made from: seedvr2_ema_7b_sharp.pth of ByteDance-Seed/SeedVR2-7B, revision eb0c4281d41ba3767d4f14370f0e37e9e9180c16
    https://huggingface.co/ByteDance-Seed/SeedVR2-7B/resolve/eb0c4281d41ba3767d4f14370f0e37e9e9180c16/seedvr2_ema_7b_sharp.pth
    SHA-256: ced5706c976d5879efcab9e108349d67abcbd8a9b36a1f48bf0f19c24164a264
  SeedVR2: Copyright (c) 2025 Bytedance Ltd. and/or its affiliates, under the Apache License, Version 2.0 (LICENSE).
  Changed: the 288 attention and MLP matrices of the blocks in int8 with comfy-kitchen's rotation (each 256 columns times the regular Hadamard / 16) and one float32 scale per row (max|row| / 127), rounded to nearest, ties to even (comfy-kitchen's INT8 layout, <layer>.weight and <layer>.weight_scale, marked by <layer>.comfy_quant); every other tensor rounded to the nearest float16.

seedvr2x_ema_7b_sharp_nvfp4.safetensors
  Made from: seedvr2_ema_7b_sharp.pth of ByteDance-Seed/SeedVR2-7B, revision eb0c4281d41ba3767d4f14370f0e37e9e9180c16
    https://huggingface.co/ByteDance-Seed/SeedVR2-7B/resolve/eb0c4281d41ba3767d4f14370f0e37e9e9180c16/seedvr2_ema_7b_sharp.pth
    SHA-256: ced5706c976d5879efcab9e108349d67abcbd8a9b36a1f48bf0f19c24164a264
  SeedVR2: Copyright (c) 2025 Bytedance Ltd. and/or its affiliates, under the Apache License, Version 2.0 (LICENSE).
  Changed: the 288 attention and MLP matrices of the blocks in NVFP4: E2M1 values, an E4M3 scale per 16 values along the input and a float32 scale per tensor (max|W| / (448 x 6)), each block's scale the one, of its max's (max / 6, to the nearest E4M3) and the 7 E4M3 values below it, that minimises the block's squared error, values rounded to nearest, ties to even (comfy-kitchen's NVFP4 layout, <layer>.weight, <layer>.weight_scale and <layer>.weight_scale_2, marked by <layer>.comfy_quant); every other tensor rounded to the nearest float16.

seedvr2x_ema_vae_fp16.safetensors
  Made from: ema_vae.pth of ByteDance-Seed/SeedVR2-7B, revision eb0c4281d41ba3767d4f14370f0e37e9e9180c16
    https://huggingface.co/ByteDance-Seed/SeedVR2-7B/resolve/eb0c4281d41ba3767d4f14370f0e37e9e9180c16/ema_vae.pth
    SHA-256: c7df8a67e68b7f9aca3d5d2153d2ce8ab4373687741a0f9ce87cb356ace51cac
  SeedVR2: Copyright (c) 2025 Bytedance Ltd. and/or its affiliates, under the Apache License, Version 2.0 (LICENSE).
  Changed: every tensor rounded from float32 to the nearest float16, ties to even; names and shapes unchanged.

transnetv2.safetensors
  Made from: TransNetV2's TensorFlow weights (inference/transnetv2-weights) of soCzech/TransNetV2, commit 85cef72af9a916bdfd7cc94a670c9cdfbf12d1ed
    https://github.com/soCzech/TransNetV2/tree/85cef72af9a916bdfd7cc94a670c9cdfbf12d1ed/inference/transnetv2-weights
    SHA-256: saved_model.pb 8ac2a52c5719690d512805b6eaf5ce12097c1d8860b3d9de245dcbbc3100f554, variables/variables.data-00000-of-00001 b8c9dc3eb807583e6215cabee9ca61737b3eb1bceff68418b43bf71459669367, variables/variables.index 8b99e28b4ad11372d9a1ad9703298c2e370df14859da4245fdbe818e92dd403f
  TransNetV2: Copyright (c) 2020 Tomáš Souček, under the MIT License (transnetv2.LICENSE).
  Changed: converted to PyTorch by TransNetV2's own inference-pytorch/convert_weights.py (its output transnetv2-pytorch-weights.pth, sha256 eed5336d5d6a013c67f5863505a26e7e835053e64a9ce413d6b089ccba07bb53), saved as safetensors; values unchanged.

seedvr2_ema_7b_fp16.imatrix.safetensors
  Made from: 4 runs of seedvr2_ema_7b_fp16.safetensors (16,479,334,424 bytes, numz's float16 file) in numz's SeedVR2 at 4490bd1f482e026674543386bb2a4d176da245b9, on 4 calibration clips: cel4k-detail.d1x4.lr.mkv, cel4k-flat.d1x4.lr.mkv, ouatia-street.d1x4.lr.mkv, ouatia-dark.d1x4.lr.mkv
  SeedVR2: Copyright (c) 2025 Bytedance Ltd. and/or its affiliates, under the Apache License, Version 2.0 (LICENSE).
  What it is: an importance matrix, not a model: for each of the 288 attention and MLP matrices of the blocks, per input channel, the sum over the runs' tokens of the input's square (in_sum2, float64) and the count of tokens (counts); in_sum2 / counts, each channel's mean square, is llama.cpp's imatrix. Collected by seedvr2x's models/gpu/imatrix_hook.py; models/seedvr2_gguf_dyn.py made seedvr2x_ema_7b_Q4_K_imatrix.gguf and seedvr2x_ema_7b_dyn.gguf with it, whose metadata name it by this name and its SHA-256.

seedvr2_ema_7b_sharp_fp16.imatrix.safetensors
  Made from: 4 runs of seedvr2_ema_7b_sharp_fp16.safetensors (16,479,334,424 bytes, numz's float16 file) in numz's SeedVR2 at 4490bd1f482e026674543386bb2a4d176da245b9, on 4 calibration clips: cel4k-detail.d1x4.lr.mkv, cel4k-flat.d1x4.lr.mkv, ouatia-street.d1x4.lr.mkv, ouatia-dark.d1x4.lr.mkv
  SeedVR2: Copyright (c) 2025 Bytedance Ltd. and/or its affiliates, under the Apache License, Version 2.0 (LICENSE).
  What it is: an importance matrix, not a model: for each of the 288 attention and MLP matrices of the blocks, per input channel, the sum over the runs' tokens of the input's square (in_sum2, float64) and the count of tokens (counts); in_sum2 / counts, each channel's mean square, is llama.cpp's imatrix. Collected by seedvr2x's models/gpu/imatrix_hook.py; models/seedvr2_gguf_dyn.py made seedvr2x_ema_7b_sharp_Q4_K_imatrix.gguf and seedvr2x_ema_7b_sharp_dyn.gguf with it, whose metadata name it by this name and its SHA-256.

	seedvr2x model files
	Copyright 2026 Edouard Hur

	These files are unofficial conversions made for seedvr2x, not their authors' releases. Each
	entry gives a file's origin, its copyright and licence, and what was changed.

	The GGUF files were quantized by ggml's own quantizer (ggml / llama.cpp, MIT License,
	Copyright (c) 2023-2026 The ggml authors), built from llama.cpp at abeada3 and called
	unchanged; no ggml code is in the files.

	seedvr2x_ema_7b_Q4_K.gguf
	Made from: seedvr2_ema_7b.pth of ByteDance-Seed/SeedVR2-7B, revision eb0c4281d41ba3767d4f14370f0e37e9e9180c16
	https://huggingface.co/ByteDance-Seed/SeedVR2-7B/resolve/eb0c4281d41ba3767d4f14370f0e37e9e9180c16/seedvr2_ema_7b.pth
	SHA-256: e1b2ae25505607e61f2a7dc7967ba778aaf3e3626d9969ce6e24c52d9ddebfcd
	SeedVR2: Copyright (c) 2025 Bytedance Ltd. and/or its affiliates, under the Apache License, Version 2.0 (LICENSE).
	Changed: the 288 attention and MLP matrices of the blocks quantized to Q4_K by ggml_quantize_chunk (llama.cpp abeada335e2e78bd3fe63febafab7e900ce75810, no importance matrix) from the float32 master; every other tensor rounded to the nearest float16.

	seedvr2x_ema_7b_Q4_K_imatrix.gguf
	Made from: seedvr2_ema_7b.pth of ByteDance-Seed/SeedVR2-7B, revision eb0c4281d41ba3767d4f14370f0e37e9e9180c16
	https://huggingface.co/ByteDance-Seed/SeedVR2-7B/resolve/eb0c4281d41ba3767d4f14370f0e37e9e9180c16/seedvr2_ema_7b.pth
	SHA-256: e1b2ae25505607e61f2a7dc7967ba778aaf3e3626d9969ce6e24c52d9ddebfcd
	SeedVR2: Copyright (c) 2025 Bytedance Ltd. and/or its affiliates, under the Apache License, Version 2.0 (LICENSE).
	Changed: the 288 attention and MLP matrices of the blocks quantized by ggml_quantize_chunk (llama.cpp abeada335e2e78bd3fe63febafab7e900ce75810) with an importance matrix (each input channel's mean square over the calibration runs of seedvr2_ema_7b_fp16.safetensors in numz's SeedVR2) from the float32 master, to Q4_K; every other tensor rounded to the nearest float16.

	seedvr2x_ema_7b_Q8_0.gguf
	Made from: seedvr2_ema_7b.pth of ByteDance-Seed/SeedVR2-7B, revision eb0c4281d41ba3767d4f14370f0e37e9e9180c16
	https://huggingface.co/ByteDance-Seed/SeedVR2-7B/resolve/eb0c4281d41ba3767d4f14370f0e37e9e9180c16/seedvr2_ema_7b.pth
	SHA-256: e1b2ae25505607e61f2a7dc7967ba778aaf3e3626d9969ce6e24c52d9ddebfcd
	SeedVR2: Copyright (c) 2025 Bytedance Ltd. and/or its affiliates, under the Apache License, Version 2.0 (LICENSE).
	Changed: the 288 attention and MLP matrices of the blocks quantized to Q8_0 by ggml_quantize_chunk (llama.cpp abeada335e2e78bd3fe63febafab7e900ce75810, no importance matrix) from the float32 master; every other tensor rounded to the nearest float16.

	seedvr2x_ema_7b_dyn.gguf
	Made from: seedvr2_ema_7b.pth of ByteDance-Seed/SeedVR2-7B, revision eb0c4281d41ba3767d4f14370f0e37e9e9180c16
	https://huggingface.co/ByteDance-Seed/SeedVR2-7B/resolve/eb0c4281d41ba3767d4f14370f0e37e9e9180c16/seedvr2_ema_7b.pth
	SHA-256: e1b2ae25505607e61f2a7dc7967ba778aaf3e3626d9969ce6e24c52d9ddebfcd
	SeedVR2: Copyright (c) 2025 Bytedance Ltd. and/or its affiliates, under the Apache License, Version 2.0 (LICENSE).
	Changed: the 288 attention and MLP matrices of the blocks quantized by ggml_quantize_chunk (llama.cpp abeada335e2e78bd3fe63febafab7e900ce75810) with an importance matrix (each input channel's mean square over the calibration runs of seedvr2_ema_7b_fp16.safetensors in numz's SeedVR2) from the float32 master, each to the type among Q3_K, Q4_K, Q5_K, Q6_K, Q8_0 that minimises the summed activation-weighted squared relative error of the 288 under the bytes of the uniform Q4_K; every other tensor rounded to the nearest float16.

	seedvr2x_ema_7b_fp16.safetensors
	Made from: seedvr2_ema_7b.pth of ByteDance-Seed/SeedVR2-7B, revision eb0c4281d41ba3767d4f14370f0e37e9e9180c16
	https://huggingface.co/ByteDance-Seed/SeedVR2-7B/resolve/eb0c4281d41ba3767d4f14370f0e37e9e9180c16/seedvr2_ema_7b.pth
	SHA-256: e1b2ae25505607e61f2a7dc7967ba778aaf3e3626d9969ce6e24c52d9ddebfcd
	SeedVR2: Copyright (c) 2025 Bytedance Ltd. and/or its affiliates, under the Apache License, Version 2.0 (LICENSE).
	Changed: every tensor rounded from float32 to the nearest float16, ties to even; names and shapes unchanged.

	seedvr2x_ema_7b_fp8_scaled.safetensors
	Made from: seedvr2_ema_7b.pth of ByteDance-Seed/SeedVR2-7B, revision eb0c4281d41ba3767d4f14370f0e37e9e9180c16
	https://huggingface.co/ByteDance-Seed/SeedVR2-7B/resolve/eb0c4281d41ba3767d4f14370f0e37e9e9180c16/seedvr2_ema_7b.pth
	SHA-256: e1b2ae25505607e61f2a7dc7967ba778aaf3e3626d9969ce6e24c52d9ddebfcd
	SeedVR2: Copyright (c) 2025 Bytedance Ltd. and/or its affiliates, under the Apache License, Version 2.0 (LICENSE).
	Changed: the 288 attention and MLP matrices of the blocks in float8_e4m3fn with one float32 scale per tensor (max\|W\| / 448; comfy-kitchen's FP8 layout, <layer>.weight and <layer>.weight_scale, marked by <layer>.comfy_quant), rounded to nearest, ties to even; every other tensor rounded to the nearest float16.

	seedvr2x_ema_7b_int8_convrot.safetensors
	Made from: seedvr2_ema_7b.pth of ByteDance-Seed/SeedVR2-7B, revision eb0c4281d41ba3767d4f14370f0e37e9e9180c16
	https://huggingface.co/ByteDance-Seed/SeedVR2-7B/resolve/eb0c4281d41ba3767d4f14370f0e37e9e9180c16/seedvr2_ema_7b.pth
	SHA-256: e1b2ae25505607e61f2a7dc7967ba778aaf3e3626d9969ce6e24c52d9ddebfcd
	SeedVR2: Copyright (c) 2025 Bytedance Ltd. and/or its affiliates, under the Apache License, Version 2.0 (LICENSE).
	Changed: the 288 attention and MLP matrices of the blocks in int8 with comfy-kitchen's rotation (each 256 columns times the regular Hadamard / 16) and one float32 scale per row (max\|row\| / 127), rounded to nearest, ties to even (comfy-kitchen's INT8 layout, <layer>.weight and <layer>.weight_scale, marked by <layer>.comfy_quant); every other tensor rounded to the nearest float16.

	seedvr2x_ema_7b_nvfp4.safetensors
	Made from: seedvr2_ema_7b.pth of ByteDance-Seed/SeedVR2-7B, revision eb0c4281d41ba3767d4f14370f0e37e9e9180c16
	https://huggingface.co/ByteDance-Seed/SeedVR2-7B/resolve/eb0c4281d41ba3767d4f14370f0e37e9e9180c16/seedvr2_ema_7b.pth
	SHA-256: e1b2ae25505607e61f2a7dc7967ba778aaf3e3626d9969ce6e24c52d9ddebfcd
	SeedVR2: Copyright (c) 2025 Bytedance Ltd. and/or its affiliates, under the Apache License, Version 2.0 (LICENSE).
	Changed: the 288 attention and MLP matrices of the blocks in NVFP4: E2M1 values, an E4M3 scale per 16 values along the input and a float32 scale per tensor (max\|W\| / (448 x 6)), each block's scale the one, of its max's (max / 6, to the nearest E4M3) and the 7 E4M3 values below it, that minimises the block's squared error, values rounded to nearest, ties to even (comfy-kitchen's NVFP4 layout, <layer>.weight, <layer>.weight_scale and <layer>.weight_scale_2, marked by <layer>.comfy_quant); every other tensor rounded to the nearest float16.

	seedvr2x_ema_7b_sharp_Q4_K.gguf
	Made from: seedvr2_ema_7b_sharp.pth of ByteDance-Seed/SeedVR2-7B, revision eb0c4281d41ba3767d4f14370f0e37e9e9180c16
	https://huggingface.co/ByteDance-Seed/SeedVR2-7B/resolve/eb0c4281d41ba3767d4f14370f0e37e9e9180c16/seedvr2_ema_7b_sharp.pth
	SHA-256: ced5706c976d5879efcab9e108349d67abcbd8a9b36a1f48bf0f19c24164a264
	SeedVR2: Copyright (c) 2025 Bytedance Ltd. and/or its affiliates, under the Apache License, Version 2.0 (LICENSE).
	Changed: the 288 attention and MLP matrices of the blocks quantized to Q4_K by ggml_quantize_chunk (llama.cpp abeada335e2e78bd3fe63febafab7e900ce75810, no importance matrix) from the float32 master; every other tensor rounded to the nearest float16.

	seedvr2x_ema_7b_sharp_Q4_K_imatrix.gguf
	Made from: seedvr2_ema_7b_sharp.pth of ByteDance-Seed/SeedVR2-7B, revision eb0c4281d41ba3767d4f14370f0e37e9e9180c16
	https://huggingface.co/ByteDance-Seed/SeedVR2-7B/resolve/eb0c4281d41ba3767d4f14370f0e37e9e9180c16/seedvr2_ema_7b_sharp.pth
	SHA-256: ced5706c976d5879efcab9e108349d67abcbd8a9b36a1f48bf0f19c24164a264
	SeedVR2: Copyright (c) 2025 Bytedance Ltd. and/or its affiliates, under the Apache License, Version 2.0 (LICENSE).
	Changed: the 288 attention and MLP matrices of the blocks quantized by ggml_quantize_chunk (llama.cpp abeada335e2e78bd3fe63febafab7e900ce75810) with an importance matrix (each input channel's mean square over the calibration runs of seedvr2_ema_7b_sharp_fp16.safetensors in numz's SeedVR2) from the float32 master, to Q4_K; every other tensor rounded to the nearest float16.

	seedvr2x_ema_7b_sharp_Q8_0.gguf
	Made from: seedvr2_ema_7b_sharp.pth of ByteDance-Seed/SeedVR2-7B, revision eb0c4281d41ba3767d4f14370f0e37e9e9180c16
	https://huggingface.co/ByteDance-Seed/SeedVR2-7B/resolve/eb0c4281d41ba3767d4f14370f0e37e9e9180c16/seedvr2_ema_7b_sharp.pth
	SHA-256: ced5706c976d5879efcab9e108349d67abcbd8a9b36a1f48bf0f19c24164a264
	SeedVR2: Copyright (c) 2025 Bytedance Ltd. and/or its affiliates, under the Apache License, Version 2.0 (LICENSE).
	Changed: the 288 attention and MLP matrices of the blocks quantized to Q8_0 by ggml_quantize_chunk (llama.cpp abeada335e2e78bd3fe63febafab7e900ce75810, no importance matrix) from the float32 master; every other tensor rounded to the nearest float16.

	seedvr2x_ema_7b_sharp_dyn.gguf
	Made from: seedvr2_ema_7b_sharp.pth of ByteDance-Seed/SeedVR2-7B, revision eb0c4281d41ba3767d4f14370f0e37e9e9180c16
	https://huggingface.co/ByteDance-Seed/SeedVR2-7B/resolve/eb0c4281d41ba3767d4f14370f0e37e9e9180c16/seedvr2_ema_7b_sharp.pth
	SHA-256: ced5706c976d5879efcab9e108349d67abcbd8a9b36a1f48bf0f19c24164a264
	SeedVR2: Copyright (c) 2025 Bytedance Ltd. and/or its affiliates, under the Apache License, Version 2.0 (LICENSE).
	Changed: the 288 attention and MLP matrices of the blocks quantized by ggml_quantize_chunk (llama.cpp abeada335e2e78bd3fe63febafab7e900ce75810) with an importance matrix (each input channel's mean square over the calibration runs of seedvr2_ema_7b_sharp_fp16.safetensors in numz's SeedVR2) from the float32 master, each to the type among Q3_K, Q4_K, Q5_K, Q6_K, Q8_0 that minimises the summed activation-weighted squared relative error of the 288 under the bytes of the uniform Q4_K; every other tensor rounded to the nearest float16.

	seedvr2x_ema_7b_sharp_fp16.safetensors
	Made from: seedvr2_ema_7b_sharp.pth of ByteDance-Seed/SeedVR2-7B, revision eb0c4281d41ba3767d4f14370f0e37e9e9180c16
	https://huggingface.co/ByteDance-Seed/SeedVR2-7B/resolve/eb0c4281d41ba3767d4f14370f0e37e9e9180c16/seedvr2_ema_7b_sharp.pth
	SHA-256: ced5706c976d5879efcab9e108349d67abcbd8a9b36a1f48bf0f19c24164a264
	SeedVR2: Copyright (c) 2025 Bytedance Ltd. and/or its affiliates, under the Apache License, Version 2.0 (LICENSE).
	Changed: every tensor rounded from float32 to the nearest float16, ties to even; names and shapes unchanged.

	seedvr2x_ema_7b_sharp_fp8_scaled.safetensors
	Made from: seedvr2_ema_7b_sharp.pth of ByteDance-Seed/SeedVR2-7B, revision eb0c4281d41ba3767d4f14370f0e37e9e9180c16
	https://huggingface.co/ByteDance-Seed/SeedVR2-7B/resolve/eb0c4281d41ba3767d4f14370f0e37e9e9180c16/seedvr2_ema_7b_sharp.pth
	SHA-256: ced5706c976d5879efcab9e108349d67abcbd8a9b36a1f48bf0f19c24164a264
	SeedVR2: Copyright (c) 2025 Bytedance Ltd. and/or its affiliates, under the Apache License, Version 2.0 (LICENSE).
	Changed: the 288 attention and MLP matrices of the blocks in float8_e4m3fn with one float32 scale per tensor (max\|W\| / 448; comfy-kitchen's FP8 layout, <layer>.weight and <layer>.weight_scale, marked by <layer>.comfy_quant), rounded to nearest, ties to even; every other tensor rounded to the nearest float16.

	seedvr2x_ema_7b_sharp_int8_convrot.safetensors
	Made from: seedvr2_ema_7b_sharp.pth of ByteDance-Seed/SeedVR2-7B, revision eb0c4281d41ba3767d4f14370f0e37e9e9180c16
	https://huggingface.co/ByteDance-Seed/SeedVR2-7B/resolve/eb0c4281d41ba3767d4f14370f0e37e9e9180c16/seedvr2_ema_7b_sharp.pth
	SHA-256: ced5706c976d5879efcab9e108349d67abcbd8a9b36a1f48bf0f19c24164a264
	SeedVR2: Copyright (c) 2025 Bytedance Ltd. and/or its affiliates, under the Apache License, Version 2.0 (LICENSE).
	Changed: the 288 attention and MLP matrices of the blocks in int8 with comfy-kitchen's rotation (each 256 columns times the regular Hadamard / 16) and one float32 scale per row (max\|row\| / 127), rounded to nearest, ties to even (comfy-kitchen's INT8 layout, <layer>.weight and <layer>.weight_scale, marked by <layer>.comfy_quant); every other tensor rounded to the nearest float16.

	seedvr2x_ema_7b_sharp_nvfp4.safetensors
	Made from: seedvr2_ema_7b_sharp.pth of ByteDance-Seed/SeedVR2-7B, revision eb0c4281d41ba3767d4f14370f0e37e9e9180c16
	https://huggingface.co/ByteDance-Seed/SeedVR2-7B/resolve/eb0c4281d41ba3767d4f14370f0e37e9e9180c16/seedvr2_ema_7b_sharp.pth
	SHA-256: ced5706c976d5879efcab9e108349d67abcbd8a9b36a1f48bf0f19c24164a264
	SeedVR2: Copyright (c) 2025 Bytedance Ltd. and/or its affiliates, under the Apache License, Version 2.0 (LICENSE).
	Changed: the 288 attention and MLP matrices of the blocks in NVFP4: E2M1 values, an E4M3 scale per 16 values along the input and a float32 scale per tensor (max\|W\| / (448 x 6)), each block's scale the one, of its max's (max / 6, to the nearest E4M3) and the 7 E4M3 values below it, that minimises the block's squared error, values rounded to nearest, ties to even (comfy-kitchen's NVFP4 layout, <layer>.weight, <layer>.weight_scale and <layer>.weight_scale_2, marked by <layer>.comfy_quant); every other tensor rounded to the nearest float16.

	seedvr2x_ema_vae_fp16.safetensors
	Made from: ema_vae.pth of ByteDance-Seed/SeedVR2-7B, revision eb0c4281d41ba3767d4f14370f0e37e9e9180c16
	https://huggingface.co/ByteDance-Seed/SeedVR2-7B/resolve/eb0c4281d41ba3767d4f14370f0e37e9e9180c16/ema_vae.pth
	SHA-256: c7df8a67e68b7f9aca3d5d2153d2ce8ab4373687741a0f9ce87cb356ace51cac
	SeedVR2: Copyright (c) 2025 Bytedance Ltd. and/or its affiliates, under the Apache License, Version 2.0 (LICENSE).
	Changed: every tensor rounded from float32 to the nearest float16, ties to even; names and shapes unchanged.

	transnetv2.safetensors
	Made from: TransNetV2's TensorFlow weights (inference/transnetv2-weights) of soCzech/TransNetV2, commit 85cef72af9a916bdfd7cc94a670c9cdfbf12d1ed
	https://github.com/soCzech/TransNetV2/tree/85cef72af9a916bdfd7cc94a670c9cdfbf12d1ed/inference/transnetv2-weights
	SHA-256: saved_model.pb 8ac2a52c5719690d512805b6eaf5ce12097c1d8860b3d9de245dcbbc3100f554, variables/variables.data-00000-of-00001 b8c9dc3eb807583e6215cabee9ca61737b3eb1bceff68418b43bf71459669367, variables/variables.index 8b99e28b4ad11372d9a1ad9703298c2e370df14859da4245fdbe818e92dd403f
	TransNetV2: Copyright (c) 2020 Tomáš Souček, under the MIT License (transnetv2.LICENSE).
	Changed: converted to PyTorch by TransNetV2's own inference-pytorch/convert_weights.py (its output transnetv2-pytorch-weights.pth, sha256 eed5336d5d6a013c67f5863505a26e7e835053e64a9ce413d6b089ccba07bb53), saved as safetensors; values unchanged.

	seedvr2_ema_7b_fp16.imatrix.safetensors
	Made from: 4 runs of seedvr2_ema_7b_fp16.safetensors (16,479,334,424 bytes, numz's float16 file) in numz's SeedVR2 at 4490bd1f482e026674543386bb2a4d176da245b9, on 4 calibration clips: cel4k-detail.d1x4.lr.mkv, cel4k-flat.d1x4.lr.mkv, ouatia-street.d1x4.lr.mkv, ouatia-dark.d1x4.lr.mkv
	SeedVR2: Copyright (c) 2025 Bytedance Ltd. and/or its affiliates, under the Apache License, Version 2.0 (LICENSE).
	What it is: an importance matrix, not a model: for each of the 288 attention and MLP matrices of the blocks, per input channel, the sum over the runs' tokens of the input's square (in_sum2, float64) and the count of tokens (counts); in_sum2 / counts, each channel's mean square, is llama.cpp's imatrix. Collected by seedvr2x's models/gpu/imatrix_hook.py; models/seedvr2_gguf_dyn.py made seedvr2x_ema_7b_Q4_K_imatrix.gguf and seedvr2x_ema_7b_dyn.gguf with it, whose metadata name it by this name and its SHA-256.

	seedvr2_ema_7b_sharp_fp16.imatrix.safetensors
	Made from: 4 runs of seedvr2_ema_7b_sharp_fp16.safetensors (16,479,334,424 bytes, numz's float16 file) in numz's SeedVR2 at 4490bd1f482e026674543386bb2a4d176da245b9, on 4 calibration clips: cel4k-detail.d1x4.lr.mkv, cel4k-flat.d1x4.lr.mkv, ouatia-street.d1x4.lr.mkv, ouatia-dark.d1x4.lr.mkv
	SeedVR2: Copyright (c) 2025 Bytedance Ltd. and/or its affiliates, under the Apache License, Version 2.0 (LICENSE).
	What it is: an importance matrix, not a model: for each of the 288 attention and MLP matrices of the blocks, per input channel, the sum over the runs' tokens of the input's square (in_sum2, float64) and the count of tokens (counts); in_sum2 / counts, each channel's mean square, is llama.cpp's imatrix. Collected by seedvr2x's models/gpu/imatrix_hook.py; models/seedvr2_gguf_dyn.py made seedvr2x_ema_7b_sharp_Q4_K_imatrix.gguf and seedvr2x_ema_7b_sharp_dyn.gguf with it, whose metadata name it by this name and its SHA-256.