Add joyfox/MiniMax-H3-Turbo as a fourth LoRA set

ComfyUI-layout 4-step turbo LoRA covering attention, MLP, AdaLN and
both output heads; per-key alpha folded into lora_B at load. Suggested
4 steps; H3_JOYFOX=off skips loading it.

README.md

@@ -60,6 +60,11 @@ A third, non-turbo set is [`fal/MiniMax-H3-Realism-People-LoRA`](https://hugging
 reference-tree layout as larry under a `diffusion_model.` prefix, so it folds through the same key mapping.
 `H3_REALISM=off` skips loading it.
 
+A fourth set is [`joyfox/MiniMax-H3-Turbo`](https://huggingface.co/joyfox/MiniMax-H3-Turbo) (`joyfox`, 4 steps):
+another ComfyUI-layout turbo LoRA, additionally covering the MLPs, the AdaLN projections, and both output heads
+(`video_out` -> `proj_out`, `audio_out` -> `audio_proj_out`), with a per-key `alpha` folded into `lora_B` at load.
+`H3_JOYFOX=off` skips loading it.
+
 ## AoTI-compiled blocks
 
 With `H3_AOTI=1` the 50 repeated transformer blocks run from a compiled package,
@@ -187,7 +192,8 @@ one-time `PIPE.to("cuda")` is inside the first row's 339 s and does not reappear
 | `H3_LORA_STRENGTH` | `1.0` | Scales the larry LoRA delta (sharpness/artifact trade-off). |
 | `H3_LIGHTX` | `on` | Set to `off` to skip loading the lightx2v LoRA set. |
 | `H3_REALISM` | `on` | Set to `off` to skip loading the fal realism-people LoRA set. |
-| `H3_LORA_DEFAULT` | `larry` | Which loaded LoRA set starts folded (`larry` / `lightx` / `realism`). |
+| `H3_JOYFOX` | `on` | Set to `off` to skip loading the joyfox 4-step turbo LoRA set. |
+| `H3_LORA_DEFAULT` | `larry` | Which loaded LoRA set starts folded (`larry` / `lightx` / `realism` / `joyfox`). |
 
 ## Whose GPU quota pays
 

app.py

@@ -304,8 +304,8 @@ def _fit_keyframe(image_path, current_canvas):
 
 
 def _resolve_lora(lora, use_lora) -> str:
-    """`lora` (`larry` / `lightx` / `realism` / `off`) wins; the legacy `use_lora` bool maps onto `larry` / `off`."""
-    if isinstance(lora, str) and lora in ("larry", "lightx", "realism", "off"):
+    """`lora` (`larry` / `lightx` / `realism` / `joyfox` / `off`) wins; the legacy bool maps onto `larry` / `off`."""
+    if isinstance(lora, str) and lora in ("larry", "lightx", "realism", "joyfox", "off"):
         return lora
     return "larry" if use_lora else "off"
 
@@ -423,7 +423,7 @@ def studio_config():
         # The LoRA dropdown: value -> {label, suggested steps}.
         "loras": {
             **{
-                name: {"label": spec["label"], "steps": {"larry": 6, "lightx": 4, "realism": 28}.get(name, 6)}
+                name: {"label": spec["label"], "steps": {"larry": 6, "lightx": 4, "realism": 28, "joyfox": 4}.get(name, 6)}
                 for name, spec in sets.items()
             },
             "off": {"label": "off (base model)", "steps": 28},

h3_lora.py

@@ -26,9 +26,14 @@ The two supported LoRAs ship in different layouts:
   `out_proj` on the 52 blocks), rank 16, `alpha == rank` (the card's scale 1.0), so it goes through the same
   `_larry_targets` mapping. It wants the full step count, not 4–6.
 
+* `joyfox` (`joyfox/MiniMax-H3-Turbo`) is another 4-step turbo LoRA, ComfyUI-native like realism but covering more
+  of the tree: attention, MLP, the block and final AdaLN projections, and both output heads (`final_layer.video_out`
+  -> `proj_out`, `final_layer.audio_out` -> `audio_proj_out`). Ranks are mixed (32 attention/MLP, 8 modulation/heads)
+  with a scalar `alpha` per key, so each `lora_B` is prescaled by `alpha / rank` at load (here always 1.0).
+
 `H3_LORA` selects the larry file (`off` skips loading it), `H3_LIGHTX=off` skips lightx, `H3_REALISM=off` skips
-realism, `H3_LORA_DEFAULT` picks which set starts folded, and `H3_LORA_STRENGTH` is the larry card's
-sharpness/artifact dial.
+realism, `H3_JOYFOX=off` skips joyfox, `H3_LORA_DEFAULT` picks which set starts folded, and `H3_LORA_STRENGTH` is
+the larry card's sharpness/artifact dial.
 """
 
 from __future__ import annotations
@@ -45,6 +50,8 @@ LIGHTX_FILE = os.environ.get("H3_LIGHTX_FILE", "minimax_h3_fl2v_turbo_4step_v0.1
 LIGHTX_ALPHA = 8
 REALISM_REPO = os.environ.get("H3_REALISM_REPO", "fal/MiniMax-H3-Realism-People-LoRA")
 REALISM_FILE = os.environ.get("H3_REALISM_FILE", "h3-realism-people-t2v.safetensors")
+JOYFOX_REPO = os.environ.get("H3_JOYFOX_REPO", "joyfox/MiniMax-H3-Turbo")
+JOYFOX_FILE = os.environ.get("H3_JOYFOX_FILE", "minimax_h3_fl2va_4step_lora.safetensors")
 # The card's sharpness/artifact dial for the larry LoRA: >1 against blurry ghosting/smear, <1 against grain.
 LARRY_STRENGTH = float(os.environ.get("H3_LORA_STRENGTH", "1.0"))
 DEFAULT_LORA = os.environ.get("H3_LORA_DEFAULT", "larry")
@@ -59,6 +66,7 @@ def _larry_targets(name: str, b: torch.Tensor, inner_dim: int) -> list[tuple[str
     else:
         target = name
     target = target.replace("final_layer.adaln_proj.linear", "norm_out.linear")
+    target = target.replace("final_layer.video_out", "proj_out").replace("final_layer.audio_out", "audio_proj_out")
 
     if target.endswith(".attn.qkv_proj"):
         prefix = target.removesuffix("qkv_proj")
@@ -131,6 +139,26 @@ def _load_realism(inner_dim: int) -> dict:
     }
 
 
+def _load_joyfox(inner_dim: int) -> dict:
+    from huggingface_hub import hf_hub_download
+    from safetensors.torch import load_file
+
+    lora = load_file(hf_hub_download(JOYFOX_REPO, JOYFOX_FILE))
+    bases = sorted({key.rsplit(".lora_", 1)[0].removeprefix("diffusion_model.") for key in lora if ".lora_" in key})
+    entries = []
+    for name in bases:
+        prefixed = f"diffusion_model.{name}"
+        a = lora[f"{prefixed}.lora_A.weight"]
+        # Per-key alpha (scalar tensor), unlike larry's fixed alpha == rank.
+        b = lora[f"{prefixed}.lora_B.weight"] * (lora[f"{prefixed}.alpha"].item() / a.shape[0])
+        entries.extend((key, a, b_part) for key, b_part in _larry_targets(name, b, inner_dim))
+    return {
+        "label": f"{JOYFOX_REPO}/{JOYFOX_FILE}",
+        "scale": 1.0,  # alpha / rank already folded into lora_B
+        "entries": entries,
+    }
+
+
 def _apply(entries, params, sign: float) -> None:
     for key, a, b in entries:
         param = params.get(key)
@@ -163,6 +191,8 @@ def apply_lora(transformer) -> str | None:
         sets["lightx"] = _load_lightx()
     if os.environ.get("H3_REALISM", "on").lower() not in ("", "off", "none"):
         sets["realism"] = _load_realism(inner_dim)
+    if os.environ.get("H3_JOYFOX", "on").lower() not in ("", "off", "none"):
+        sets["joyfox"] = _load_joyfox(inner_dim)
     if not sets:
         return None