#!/usr/bin/env python3 """Build h3-exp-023's 24 arms: does a SAME-SUBJECT guide video transfer action? THE OBJECTION THIS EXISTS TO ANSWER h3-exp-010 (guide -> H3FunControlApply.control_video) and h3-exp-012 (guide -> MiniMaxH3ReferenceToVideo.ref_videos) both found that a Wan2.1 VACE guide transferred neither the briefcase/tripod props nor the kneel, while an OpenPose skeleton over the same choreography transferred the action cleanly. Both used a guide depicting a suited businessman while prompting for a woman in a 1968 mini dress. A defender of the guide-video recipe can say the guide never had a fair run: conditioning on subject A while prompting subject B is a direct semantic collision, so the null result may be about the mismatch rather than about guide conditioning. That objection is legitimate and cheap to settle. THE DESIGN: 2x2 PLUS TWO CONTROLS, 4 SEEDS EACH control_video ref_videos different-subject arm1_fc_diff arm3_ref_diff <- exp-010/012 redone same-subject arm2_fc_same arm4_ref_same <- the novel cells armS_pose OpenPose skeleton -> control_video positive control: action DID transfer here before arm0_null no control video, no reference negative control: what the prompt alone produces Six conditions x seeds 43/44/45/46 = 24 renders, sequential under gpu-lock. WHAT IS HELD CONSTANT, AND ONE DEVIATION FROM THE DESIGN DOC Everything but the conditioning input: prompt, geometry, model stack, sampler, seeds. The prompt is exp-010's, byte-for-byte, asserted below -- the design doc paraphrased it as "A woman in a straight-cut dress with red hem walking in a room..." but no such fixture exists; the real held-constant prompt is the 1968 mini-dress one every arm in this arc has used. Using the doc's paraphrase would have silently changed the subject the guides are supposed to match, which is the exact confound this experiment was written to remove. Props stay OUT of the prompt on purpose: if the prompt names the briefcase, prop appearance is no longer evidence that the GUIDE carried it. Geometry is exp-014's cheap shot -- 832x480, 124 frames (17*7+5) -- not exp-010's 1312x736x294. Nothing here is compared against exp-010's numbers directly (arm1/arm3 re-run those conditions inside this experiment), so the smaller shot buys 4 seeds per arm instead of 2, and 24 runs that fit one overnight lock window. """ import copy import hashlib import json import os import sys HERE = os.path.dirname(os.path.abspath(__file__)) REPO = os.path.dirname(os.path.dirname(HERE)) SEEDS = [43, 44, 45, 46] W, H, LEN = 832, 480, 124 # 124 = 17*7 + 5 GUIDE_LEN = 90 # 90 = 17*5 + 5, inside ref_videos' window assert (LEN - 5) % 17 == 0, "shot length must satisfy 17n+5" assert (GUIDE_LEN - 5) % 17 == 0, "reference length must satisfy 17n+5" # arm -> (conditioning pathway, fixture installed by build_exp023_fixtures.py) # "fc" : H3FunControlApply.control_video, fixture must be shot-length (124) # "ref" : MiniMaxH3ReferenceToVideo.ref_videos, fixture is 90 # "none" : neither pathway wired ARMS = { "arm1_fc_diff": ("fc", "exp023_guide_diff_124.mp4"), "arm2_fc_same": ("fc", "exp023_guide_same_124.mp4"), "arm3_ref_diff": ("ref", "exp023_guide_diff_90.mp4"), "arm4_ref_same": ("ref", "exp023_guide_same_90.mp4"), "armS_pose": ("fc", "exp023_pose_124.mp4"), "arm0_null": ("none", None), } def main(): src = os.path.join(REPO, "workflows", "h3-exp-023-base", "h3_base.api.json") out = os.path.join(REPO, "workflows", "h3-exp-023-arms") os.makedirs(out, exist_ok=True) base = json.load(open(src)) # Pin the stack the design registered, so a swapped loader fails here and # not three hours into the lock. assert base["3"]["inputs"]["clip_name"] == \ "qwen3vl_32b_minimax_h3_nvfp4_awq.safetensors", "text encoder drifted (ADR 0014)" assert base["4"]["inputs"]["unet_name"] == \ "MinimaxH3/minimax_h3_ref2va_pruned_int8_convrot.safetensors", "diffusion model drifted" PROMPT = base["11"]["inputs"]["prompt"] assert "blonde bob" in PROMPT and "1968" in PROMPT, "prompt is not the exp-010 prompt" for banned in ("briefcase", "tripod", "kneel"): assert banned not in PROMPT.lower(), ( f"prompt names '{banned}' -- prop/action appearance would then be " "attributable to the text, not the guide") written = [] for arm, (pathway, fixture) in sorted(ARMS.items()): for seed in SEEDS: g = copy.deepcopy(base) g["9"]["inputs"]["noise_seed"] = seed g["11"]["inputs"]["width"] = W g["11"]["inputs"]["height"] = H g["11"]["inputs"]["length"] = LEN if pathway == "fc": g["60"]["inputs"]["file"] = fixture elif pathway == "ref": g["60"]["inputs"]["file"] = fixture # rewire: drop FunControl, feed the clip to ref_videos instead passthrough = g["63"]["inputs"]["model"] for node in g.values(): for k, v in (node.get("inputs") or {}).items(): if isinstance(v, list) and len(v) == 2 and str(v[0]) == "63": node["inputs"][k] = passthrough del g["63"] g.pop("62", None) g["11"]["inputs"]["ref_videos.ref_video_0"] = ["61", 0] else: # none passthrough = g["63"]["inputs"]["model"] for node in g.values(): for k, v in (node.get("inputs") or {}).items(): if isinstance(v, list) and len(v) == 2 and str(v[0]) == "63": node["inputs"][k] = passthrough del g["63"] for nid in ("62", "60", "61"): g.pop(nid, None) name = f"{arm}_s{seed}" g["19"]["inputs"]["filename_prefix"] = f"H3/exp023/{name}" path = os.path.join(out, f"{name}.api.json") json.dump(g, open(path, "w"), indent=2) written.append((name, path, arm, pathway, fixture)) # ---- post-conditions ------------------------------------------------- # The h3-exp-001 lesson: three "arms" that were byte-identical apart from # SaveVideo rendered identically, Immich deduped them to one asset, and the # comparison was void after the GPU time was spent. seen = {} for name, path, arm, pathway, fixture in written: g = json.load(open(path)) g.pop("19", None) h = hashlib.sha256(json.dumps(g, sort_keys=True).encode()).hexdigest()[:16] assert h not in seen, f"{name} is identical to {seen[h]}" seen[h] = name n11 = [x for x in g.values() if x.get("class_type") == "MiniMaxH3ReferenceToVideo"][0] has_ref = any(k.startswith("ref_videos") for k in n11["inputs"]) has_fc = any(x.get("class_type") == "H3FunControlApply" for x in g.values()) loader = [x for x in g.values() if x.get("class_type") == "LoadVideo"] assert has_fc == (pathway == "fc"), f"{name}: FunControl wiring wrong" assert has_ref == (pathway == "ref"), f"{name}: ref_videos wiring wrong" # Exactly one pathway, never both -- otherwise an effect is # unattributable between them. assert not (has_fc and has_ref), f"{name}: both pathways wired" if pathway == "none": assert not loader, f"{name}: null arm still loads a video" else: assert len(loader) == 1 and loader[0]["inputs"]["file"] == fixture, \ f"{name}: expected fixture {fixture}" assert n11["inputs"]["prompt"] == PROMPT, f"{name}: prompt drifted" assert (n11["inputs"]["width"], n11["inputs"]["height"], n11["inputs"]["length"]) == (W, H, LEN), f"{name}: geometry drifted" assert g["9"]["inputs"]["noise_seed"] in SEEDS, f"{name}: seed off-design" # Every arm must differ from every other arm at the SAME seed in exactly the # conditioning input, and every seed within an arm must differ only in seed. by_arm = {} for name, path, arm, _p, _f in written: by_arm.setdefault(arm, []).append(name) assert len(by_arm) == 6, f"expected 6 arms, got {sorted(by_arm)}" for arm, members in sorted(by_arm.items()): assert len(members) == len(SEEDS), f"{arm}: {len(members)} seeds" print(f" {arm:15} {len(members)} seeds pathway={ARMS[arm][0]:5} " f"fixture={ARMS[arm][1]}") print(f"\nOK: {len(written)} arms ({len(by_arm)} conditions x {len(SEEDS)} seeds), " f"all hashes distinct, prompt and geometry identical across all") return 0 if __name__ == "__main__": sys.exit(main())