Add fal/MiniMax-H3-Realism-People-LoRA as a third LoRA set

Style LoRA (trigger word r34l1sm), reference-tree keys under a
diffusion_model. prefix, folded through the same mapping as larry.
Suggested 28 steps; H3_REALISM=off skips loading it.

README.md

@@ -55,6 +55,11 @@ and the low-rank factors stay resident so switching is an in-place unfold/fold t
 `H3_LORA` selects the larry file (`off` skips it), `H3_LIGHTX=off` skips lightx, `H3_LORA_DEFAULT` picks which set
 starts folded, and `H3_LORA_STRENGTH` scales the larry update (the card's sharpness/artifact dial).
 
+A third, non-turbo set is [`fal/MiniMax-H3-Realism-People-LoRA`](https://huggingface.co/fal/MiniMax-H3-Realism-People-LoRA)
+(`realism` in the dropdown, trigger word `r34l1sm`): realistic people at the full 28 steps. It ships in the same
+reference-tree layout as larry under a `diffusion_model.` prefix, so it folds through the same key mapping.
+`H3_REALISM=off` skips loading it.
+
 ## AoTI-compiled blocks
 
 With `H3_AOTI=1` the 50 repeated transformer blocks run from a compiled package,
@@ -181,7 +186,8 @@ one-time `PIPE.to("cuda")` is inside the first row's 339 s and does not reappear
 | `H3_LORA_REPO` | `larryvrh/MiniMax-H3-Turbo-Lora` | Hub repo the LoRA is fetched from. |
 | `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_LORA_DEFAULT` | `larry` | Which loaded LoRA set starts folded (`larry` / `lightx`). |
+| `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`). |
 
 ## 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` / `off`) wins; the legacy `use_lora` bool maps onto `larry` / `off`."""
-    if isinstance(lora, str) and lora in ("larry", "lightx", "off"):
+    """`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"):
         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}.get(name, 6)}
+                name: {"label": spec["label"], "steps": {"larry": 6, "lightx": 4, "realism": 28}.get(name, 6)}
                 for name, spec in sets.items()
             },
             "off": {"label": "off (base model)", "steps": 28},

h3_lora.py

@@ -21,8 +21,14 @@ The two supported LoRAs ship in different layouts:
   `transformer_blocks.N.attn.to_q.lora_A.default.weight` and friends — rank 128, `alpha == 8`, so the fold scale is
   `8 / 128 = 0.0625` (matching `set_adapters(weights=1.0)` in their inference script). Keys map name-for-name.
 
-`H3_LORA` selects the larry file (`off` skips loading it), `H3_LIGHTX=off` skips lightx, `H3_LORA_DEFAULT` picks
-which set starts folded, and `H3_LORA_STRENGTH` is the larry card's sharpness/artifact dial.
+* `realism` (`fal/MiniMax-H3-Realism-People-LoRA`) is a *style* LoRA, not a turbo one — realistic people, trigger
+  word `r34l1sm`. Same reference tree as larry under a `diffusion_model.` prefix, attention only (`qkv_proj` /
+  `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.
+
+`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.
 """
 
 from __future__ import annotations
@@ -37,6 +43,8 @@ LARRY_FILE = os.environ.get("H3_LORA", "minimax_h3_turbo_4step_ema_ckpt850.safet
 LIGHTX_REPO = os.environ.get("H3_LIGHTX_REPO", "lightx2v/Minimax-h3-Turbo")
 LIGHTX_FILE = os.environ.get("H3_LIGHTX_FILE", "minimax_h3_fl2v_turbo_4step_v0.1.safetensors")
 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")
 # 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")
@@ -105,6 +113,24 @@ def _load_lightx() -> dict:
     }
 
 
+def _load_realism(inner_dim: int) -> dict:
+    from huggingface_hub import hf_hub_download
+    from safetensors.torch import load_file
+
+    lora = load_file(hf_hub_download(REALISM_REPO, REALISM_FILE))
+    bases = sorted({key.rsplit(".lora_", 1)[0].removeprefix("diffusion_model.") for key in lora})
+    entries = []
+    for name in bases:
+        a = lora[f"diffusion_model.{name}.lora_A.weight"]
+        b = lora[f"diffusion_model.{name}.lora_B.weight"]
+        entries.extend((key, a, b_part) for key, b_part in _larry_targets(name, b, inner_dim))
+    return {
+        "label": f"{REALISM_REPO}/{REALISM_FILE}",
+        "scale": 1.0,  # alpha == rank, per the card's scale 1.0
+        "entries": entries,
+    }
+
+
 def _apply(entries, params, sign: float) -> None:
     for key, a, b in entries:
         param = params.get(key)
@@ -135,6 +161,8 @@ def apply_lora(transformer) -> str | None:
         sets["larry"] = _load_larry(inner_dim)
     if os.environ.get("H3_LIGHTX", "on").lower() not in ("", "off", "none"):
         sets["lightx"] = _load_lightx()
+    if os.environ.get("H3_REALISM", "on").lower() not in ("", "off", "none"):
+        sets["realism"] = _load_realism(inner_dim)
     if not sets:
         return None