Add files using upload-large-folder tool

MANIFEST.json

@@ -0,0 +1,35 @@
+{
+  "schema_version": 1,
+  "files": [
+    {
+      "name": "README.md",
+      "bytes": 2855,
+      "sha256": "8c711dd06411fb869070e8ed6cee0fa02b73ec545dba7220c602eea82f3aad52"
+    },
+    {
+      "name": "conditioner.pt",
+      "bytes": 128615217,
+      "sha256": "77c38f544576c26c80e0c1767bfb02802572266c9de95baf99265d18eafda909"
+    },
+    {
+      "name": "evaluation.json",
+      "bytes": 4790,
+      "sha256": "626a5bb0a49e5f01457a2cc707c2a00a82f4359f38638a3f37e10f1ea3626130"
+    },
+    {
+      "name": "mugen_config.json",
+      "bytes": 679,
+      "sha256": "427f0e93dea1e4715bcf0f2adf710fe6a16f046bd55dfab35a4b043ae01b9bfa"
+    },
+    {
+      "name": "pytorch_lora_weights.safetensors",
+      "bytes": 11824784,
+      "sha256": "9d164a72cca4195070da8d1ccc2b42258673f4c64e88c3304eba32ea78e0fcfd"
+    }
+  ],
+  "forbidden_files_absent": [
+    "optimizer.pt",
+    "AnyFlow base weights",
+    "dataset media"
+  ]
+}

README.md

@@ -1,5 +1,65 @@
 ---
+library_name: diffusers
 license: other
 license_name: mit-code-with-noncommercial-upstream-weight-restrictions
-license_link: LICENSE
+base_model: nvidia/AnyFlow-FAR-Wan2.1-1.3B-Diffusers
+pipeline_tag: image-to-video
+tags:
+  - multimodal
+  - video-generation
+  - lora
+  - retrieval
 ---
+
+# MUGen-VR
+
+MUGen-VR adds project-owned multimodal condition modules to AnyFlow-FAR. ImageBind
+encodes text, image, and audio; InternVideo2 encodes videos and retrieved references.
+Three hierarchical fusion tokens and four score-aware reference tokens are projected
+to the 4096-dimensional UMT5 prompt space and appended to AnyFlow `prompt_embeds`.
+
+## Published files
+
+The Hugging Face release contains only MUGen-owned lightweight files:
+
+- Fusion and Reference Adapter weights.
+- The condition projector and modality/type embeddings.
+- AnyFlow cross-attention LoRA weights.
+- Training configuration, encoder/data versions, and evaluation metadata.
+
+It does not contain AnyFlow, ImageBind, InternVideo2, MSR-VTT media, or cached third-party
+features. Users must obtain those assets under their upstream terms.
+
+## Evaluation
+
+The practical project evaluation compares four variants on 40 fixed held-out samples
+with generation seed 42:
+
+Condition scale `0.1` was selected on eight validation samples before the fixed test run.
+
+| Variant | Definition | VBench | Retrieval MRR | Audio-flow | Latency |
+|---|---|---:|---:|---:|---:|
+| B0 | AnyFlow image + original prompt | 0.7600 | n/a | 0.0179 | 4.42 s |
+| B1 | prompt rewrite prototype | 0.7511 | n/a | 0.0363 | 4.26 s |
+| B2 | Fusion tokens without audio/reference | 0.7596 | 1.0000 | -0.0012 | 4.28 s |
+| B3 | full Fusion + audio + reference | 0.7570 | 0.9813 | 0.0046 | 4.31 s |
+
+The completion check passed for all 40 pairs, four variants, one seed, required metrics,
+and eight showcase cases. B0 remained the strongest VBench baseline. B2 preserved quality
+within 0.0004, while audio/reference conditioning in B3 did not improve retrieval, audio
+control, or aggregate video quality. The release therefore claims a reproducible real
+condition path and an honest negative result, not a generation-quality gain.
+
+## Limitations
+
+- The base AnyFlow model is restricted to non-commercial use under NVIDIA NSCLv1.
+- ImageBind is also governed by non-commercial research terms.
+- Audio control is indirect through condition tokens; it does not synthesize an output soundtrack.
+- Retrieval quality depends on the licensed local reference gallery.
+- On this checkpoint, adding audio and reference tokens slightly reduced aggregate VBench and retrieval MRR.
+- The project is evaluated as an internship portfolio system, not a paper-scale benchmark.
+
+## Citation
+
+Please cite the upstream AnyFlow, ImageBind, InternVideo2, MSR-VTT, and VBench projects.
+See `THIRD_PARTY_NOTICES.md` and `docs/third_party_commits.md` in the source repository.

conditioner.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:77c38f544576c26c80e0c1767bfb02802572266c9de95baf99265d18eafda909
+size 128615217

evaluation.json

@@ -0,0 +1,152 @@
+{
+  "schema_version": 1,
+  "summary": {
+    "B0": {
+      "description": "AnyFlow image + original prompt",
+      "count": 40,
+      "means": {
+        "latency_seconds": 4.415573643179959,
+        "generated_fps": 5.700837394230057,
+        "peak_vram_gib": 15.610107421875,
+        "subject_consistency": 0.884030045941472,
+        "motion_smoothness": 0.984162795787803,
+        "temporal_flickering": 0.9728844855112184,
+        "aesthetic_quality": 0.4435853444039819,
+        "imaging_quality": 0.5152302944946286,
+        "vbench_total": 0.759978593227821,
+        "imagebind_audio_video_alignment": 0.2249340817797929,
+        "onset_flow_correlation": 0.01786091947584697
+      }
+    },
+    "B1": {
+      "description": "prompt rewrite prototype",
+      "count": 40,
+      "means": {
+        "latency_seconds": 4.26074977768003,
+        "generated_fps": 5.877865403882483,
+        "peak_vram_gib": 15.610107421875,
+        "subject_consistency": 0.8433478668642538,
+        "motion_smoothness": 0.9751392628313267,
+        "temporal_flickering": 0.9570126858178301,
+        "aesthetic_quality": 0.45098427012562714,
+        "imaging_quality": 0.5291726571655279,
+        "vbench_total": 0.7511313485609136,
+        "imagebind_audio_video_alignment": 0.23401265256106854,
+        "onset_flow_correlation": 0.03631851392772683
+      }
+    },
+    "B2": {
+      "description": "fusion tokens without audio or reference",
+      "count": 40,
+      "means": {
+        "latency_seconds": 4.279900551162427,
+        "generated_fps": 5.849079877479161,
+        "peak_vram_gib": 15.610107421875,
+        "subject_consistency": 0.8709420069120819,
+        "motion_smoothness": 0.9821310122763294,
+        "temporal_flickering": 0.9698621026908654,
+        "aesthetic_quality": 0.44844568669796,
+        "imaging_quality": 0.5265446289634709,
+        "vbench_total": 0.7595850875081422,
+        "imagebind_audio_video_alignment": 0.2250766169745475,
+        "onset_flow_correlation": -0.0011641985030232075,
+        "retrieval_rank": 1.0,
+        "retrieval_mrr": 1.0
+      }
+    },
+    "B3": {
+      "description": "full fusion + audio + reference",
+      "count": 40,
+      "means": {
+        "latency_seconds": 4.309116636033286,
+        "generated_fps": 5.810298692646417,
+        "peak_vram_gib": 15.610107421875,
+        "subject_consistency": 0.8677799109912785,
+        "motion_smoothness": 0.9811141542968373,
+        "temporal_flickering": 0.9684628730077376,
+        "aesthetic_quality": 0.44271424114704105,
+        "imaging_quality": 0.524960030918121,
+        "vbench_total": 0.7570062420722037,
+        "imagebind_audio_video_alignment": 0.22060853205621242,
+        "onset_flow_correlation": 0.0046293259929009465,
+        "retrieval_rank": 1.075,
+        "retrieval_mrr": 0.98125
+      }
+    }
+  },
+  "paired_bootstrap_advisory": {
+    "generated_fps": {
+      "best_baseline": "B1",
+      "pairs": 40,
+      "delta": -0.06756671123606395,
+      "ci_lower": -0.13503826082529,
+      "ci_upper": -0.0009061307435047152,
+      "confidence": 0.95,
+      "bootstrap_samples": 10000
+    },
+    "imagebind_audio_video_alignment": {
+      "best_baseline": "B1",
+      "pairs": 40,
+      "delta": -0.01340412050485611,
+      "ci_lower": -0.027256709835492065,
+      "ci_upper": -0.000871459546033293,
+      "confidence": 0.95,
+      "bootstrap_samples": 10000
+    },
+    "latency_seconds": {
+      "best_baseline": "B1",
+      "pairs": 40,
+      "delta": -0.04836685835325625,
+      "ci_lower": -0.09987435037262912,
+      "ci_upper": 0.0023366264591459175,
+      "confidence": 0.95,
+      "bootstrap_samples": 10000
+    },
+    "onset_flow_correlation": {
+      "best_baseline": "B1",
+      "pairs": 40,
+      "delta": -0.031689187934825884,
+      "ci_lower": -0.08407993431787751,
+      "ci_upper": 0.020294409306492248,
+      "confidence": 0.95,
+      "bootstrap_samples": 10000
+    },
+    "peak_vram_gib": {
+      "best_baseline": "B0",
+      "pairs": 40,
+      "delta": 0.0,
+      "ci_lower": 0.0,
+      "ci_upper": 0.0,
+      "confidence": 0.95,
+      "bootstrap_samples": 10000
+    },
+    "retrieval_mrr": {
+      "best_baseline": "B2",
+      "pairs": 40,
+      "delta": -0.01875,
+      "ci_lower": -0.05625,
+      "ci_upper": 0.0,
+      "confidence": 0.95,
+      "bootstrap_samples": 10000
+    },
+    "retrieval_rank": {
+      "best_baseline": "B2",
+      "pairs": 40,
+      "delta": 0.075,
+      "ci_lower": 0.0,
+      "ci_upper": 0.225,
+      "confidence": 0.95,
+      "bootstrap_samples": 10000
+    }
+  },
+  "completion_check": {
+    "checks": {
+      "held_out_pairs_30_to_50": true,
+      "all_pairs_have_b0_to_b3": true,
+      "uniform_generation_seed": true,
+      "required_metrics_present": true,
+      "side_by_side_cases_6_to_10": true
+    },
+    "passed": true
+  }
+}

mugen_config.json

@@ -0,0 +1,29 @@
+{
+  "schema_version": 1,
+  "base_model": "nvidia/AnyFlow-FAR-Wan2.1-1.3B-Diffusers",
+  "checkpoint_step": 300,
+  "condition_scale": 0.1,
+  "reference_tokens": 4,
+  "reference_top_k": 3,
+  "lora": {
+    "rank": 16,
+    "alpha": 32
+  },
+  "generation": {
+    "num_frames": 25,
+    "height": 256,
+    "width": 448,
+    "latent_chunk_size": 2
+  },
+  "encoder_versions": {
+    "imagebind": "ImageBind@53680b0",
+    "internvideo": "InternVideo@3965eef:Stage2-1B-224p-f4"
+  },
+  "feature_rows": 6000,
+  "training_metrics": {
+    "train_loss": 0.23321709036827087,
+    "validation_loss": 0.4095851182937622
+  },
+  "upstream_weights_included": false,
+  "optimizer_state_included": false
+}

pytorch_lora_weights.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:9d164a72cca4195070da8d1ccc2b42258673f4c64e88c3304eba32ea78e0fcfd
+size 11824784