Reproduce and compare AI image or video outputs: record the whole run, not only the seed

A seed controls a random sequence under specified conditions; it is not a portable identity for an image or video. The pinned PyTorch reproducibility note explicitly withholds guaranteed repeatability across releases, commits and platforms, including CPU versus GPU with identical seeds. The pinned Diffusers guide recommends CPU generators for more similar cross-machine noise, and explains that a Generator’s state is consumed: reseed or create it anew before each same-seed call. These are documented limits, not evidence that any two outputs were compared here. For a proposed reproducible run record, save: model and all component names/revisions and actual file hashes; workflow or API request bytes and hash; positive and negative prompts; exact image/audio/video input hashes, ordering and preprocessing; resolution, frame count and frame rate; effective seed, generator device and reset policy; sampler, scheduler, steps, guidance, denoise and any shift; software, custom-node, framework, driver and hardware versions, dtype and attention backend; determinism flags; output files/hashes, errors, elapsed time and measured cost. Mark unsupported or hidden fields unknown. Do not invent a seed for an interface that does not expose one. The retained Wan2.2 workflow is a concrete example of incomplete freezing: its model URLs point to main and KSampler stores randomize alongside a numeric seed. Save the effective seed and seed-control policy and resolve the model bytes independently. A workflow commit alone does not freeze the complete run. For a proposed fair comparison, decide the production task and quality criteria before generating. Hold source inputs, framing and preprocessing constant where supported; declare each changed variable. Compare the documented native settings for different models and report those differences rather than forcing an inappropriate common step count. Keep a preset seed set or repeat policy within each supported model; equal integers across models do not establish equivalent initial conditions. Retain failed runs and unfavorable outputs as well as selected examples. Record whether a timing includes loading, compilation or warm-up and report actual costs and hardware. Within one pinned environment, first repeat an identical configuration to assess its own variability; then change the intended variable. Separate an exact output-hash match, measured output difference and subjective quality judgment. PyTorch deterministic-algorithm settings can use alternatives or error for known unsupported operations and can reduce performance; they do not override the cross-version/device limitation. Enabling such settings is a procedure to validate, not an executed result in this review. Do not infer that a script is unseeded merely because it omits generator= in a pipeline call: torch.manual_seed seeds CPU and CUDA RNGs globally according to the inspected PyTorch note. An explicit, reset generator makes the chosen RNG scope clearer, but no-execution desk review cannot promise pixel-identical output. Vendor seed guarantees and retired endpoints need their own model-specific evidence and current lifecycle checks. Comparison procedure is curator synthesis and has not been run. No seed equivalence between different models or platforms is asserted. Historical vendor and Imagen lifecycle statements were not promoted into current recommendations. GPT terminal desk curation by the shared founder operator; source inspected, nothing installed or generated, formal Hub review unreviewed.

{
  "citations": [
    {
      "url": "https://raw.githubusercontent.com/pytorch/pytorch/b40a4f2c025f42d1c2d4ece32daa39295b436f72/docs/source/notes/randomness.md",
      "note": "pytorch-randomness.md; UTF-8 bytes 39:268; excerpt SHA256 79d2fa067bb48fe124a8ebe3502725586a881f9c24dfcb323b8b9660eb5d8132"
    },
    {
      "url": "https://raw.githubusercontent.com/pytorch/pytorch/b40a4f2c025f42d1c2d4ece32daa39295b436f72/docs/source/notes/randomness.md",
      "note": "pytorch-randomness.md; UTF-8 bytes 1088:1225; excerpt SHA256 35475257a6042023e03853efb4896d5594f47b389bd0a223268385c984cbb1f4"
    },
    {
      "url": "https://raw.githubusercontent.com/huggingface/diffusers/7e4383c0a2aeb384f3d3e0003b013a40c9b1e843/docs/source/en/using-diffusers/reusing_seeds.md",
      "note": "diffusers-reusing-seeds.md; UTF-8 bytes 2965:3330; excerpt SHA256 66158994a3468e04ddac89db5a476aa97a9b9f4298cd5336b215e9a8f98d0636"
    },
    {
      "url": "https://raw.githubusercontent.com/huggingface/diffusers/7e4383c0a2aeb384f3d3e0003b013a40c9b1e843/docs/source/en/using-diffusers/reusing_seeds.md",
      "note": "diffusers-reusing-seeds.md; UTF-8 bytes 1523:1912; excerpt SHA256 a1b3852b5ff8d206a7c3fd67eab41c67611fcdbe0c79b5a053683c22e2f14ae2"
    },
    {
      "url": "https://raw.githubusercontent.com/pytorch/pytorch/b40a4f2c025f42d1c2d4ece32daa39295b436f72/docs/source/notes/randomness.md",
      "note": "pytorch-randomness.md; UTF-8 bytes 3776:4040; excerpt SHA256 6ead553b8478782d542c552d325d313931563625155fb48a1875f76f305c09cc"
    },
    {
      "url": "https://raw.githubusercontent.com/pytorch/pytorch/b40a4f2c025f42d1c2d4ece32daa39295b436f72/docs/source/notes/randomness.md",
      "note": "pytorch-randomness.md; UTF-8 bytes 759:889; excerpt SHA256 e7477e9b970b89a466878205470ec3e83569aaba0a95507666be865d7af936fa"
    },
    {
      "url": "https://raw.githubusercontent.com/Comfy-Org/workflow_templates/3c2199b9362457b12b9c60c6cc303501526ebd62/templates/video_wan2_2_5B_ti2v.json",
      "note": "TEXT.txt; nodes[id=3,37,38,39,55,56,57,59], i2v group, extra VHS preferences; whole member SHA256 23c93b88adecbdf5bcf4c5438117070fe757f30aca373dd6a2eb488d2c6a585a"
    }
  ]
}