lora: trainers for LTX-2, MiniMax-H3 and Krea 2, selectable per request

We had trainers for SD1.x/SDXL, Z-Image and Wan — none for the architectures the
LoRAs people actually publish for video today. All three new ones are the same
training problem as Wan (encode references to VAE latents, encode the prompt once,
adapt the transformer's attention projections, regress the rectified-flow
velocity), so rather than a fourth, fifth and sixth copy of that loop there is one
shared _train_flow_dit driven by a per-architecture description in
codai/api/lora_archs.py: which classes to import, the latent shape, the timestep
shift, the text-encoder contract.

Selection: `target` accepts image | video | wan | ltx2 | h3 | krea. `video` keeps
meaning Wan for existing configs but now reads the model id, so an LTX-2 or H3 path
routes itself; the explicit names force one. The character studio exposes the same
choice as --target and a dropdown.

Availability differs per architecture and fails with a message that says so rather
than an ImportError from inside a training job: LTX-2 works on this diffusers
(0.38 has LTX2VideoTransformer3DModel + AutoencoderKLLTX2Video), while H3 and
Krea 2 need >= 0.40 — and Krea 2 additionally needs access to the gated
krea/Krea-2-* weights.

Verified: architecture detection and routing for every target/model combination,
and that each trainer's classes either load or explain what they need. NOT run
against real weights for any of the three — the contracts match the proven Wan
path, but the first real run should be treated as a debugging session.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_0182DhR1eFNPQyGmDrpHbedY

codai/api/lora_archs.py

+# CoderAI - OpenAI-compatible API server
+# Copyright (C) 2026 Stefy Lanza <stefy@nexlab.net>
+#
+# This program is free software: you can redistribute it and/or modify
+# it under the terms of the GNU General Public License as published by
+# the Free Software Foundation, either version 3 of the License, or
+# (at your option) any later version.
+
+"""Which architecture a LoRA is being trained for, and what that implies.
+
+`codai/api/loras.py` grew one trainer per architecture (SD1.x, SDXL, Z-Image,
+Wan). The newer targets — LTX-2, MiniMax-H3, Krea 2 — are all the same shape as
+Wan: a flow-matching transformer over VAE latents, conditioned on a text
+embedding. So rather than a fourth, fifth and sixth copy of that loop, each one is
+described here and driven by the shared trainer.
+
+Each entry says what to import, what the latents look like, and how strongly to
+shift the rectified-flow timestep. Nothing here imports diffusers at module level:
+an architecture whose classes this diffusers build doesn't have must fail with a
+clear message at TRAIN time, not stop the server from starting.
+"""
+
+from typing import Optional
+
+#: target -> description. `detect` matches a model path/id (lowercased).
+ARCHS = {
+    "wan": {
+        "label": "Wan video DiT",
+        "kind": "video",
+        "detect": ("wan2.1", "wan2.2", "wan-t2v", "wan-i2v", "wanvace", "wan_"),
+        "transformer": "WanTransformer3DModel",
+        "vae": "AutoencoderKLWan",
+        "text_encoder": ("transformers", "UMT5EncoderModel"),
+        "shift": 5.0,
+        "note": "Handled by the dedicated _train_wan (dual-expert aware).",
+    },
+    "ltx2": {
+        "label": "LTX-2 video DiT",
+        "kind": "video",
+        "detect": ("ltx-2", "ltx2", "ltx-video-2"),
+        "transformer": "LTX2VideoTransformer3DModel",
+        "vae": "AutoencoderKLLTX2Video",
+        "text_encoder": ("transformers", "T5EncoderModel"),
+        "tokenizer": ("transformers", "AutoTokenizer"),
+        "max_text_len": 256,
+        "shift": 3.0,
+    },
+    "h3": {
+        "label": "MiniMax-H3 omni DiT",
+        "kind": "video",
+        "detect": ("minimax-h3", "minimax_h3"),
+        "transformer": "MiniMaxH3Transformer3DModel",
+        "vae": "AutoencoderKLMiniMaxH3",
+        "text_encoder": ("transformers", "AutoModel"),
+        "tokenizer": ("transformers", "AutoTokenizer"),
+        "max_text_len": 512,
+        # H3 carries separate video/audio flow shifts (12 / 3); training stills we
+        # only drive the video branch.
+        "shift": 12.0,
+        "needs": "diffusers>=0.40",
+    },
+    "krea": {
+        "label": "Krea 2 image DiT",
+        "kind": "image",
+        "detect": ("krea-2", "krea2", "krea_2"),
+        "transformer": "Krea2Transformer2DModel",
+        "vae": "AutoencoderKL",
+        "text_encoder": ("transformers", "AutoModel"),
+        "tokenizer": ("transformers", "AutoTokenizer"),
+        "max_text_len": 512,
+        "shift": 3.0,
+        "needs": "diffusers>=0.40 and access to the gated krea/Krea-2-* weights",
+    },
+}
+
+#: What `target` may be on a train request.
+TARGETS = ("image", "video", "wan", "ltx2", "h3", "krea")
+
+
+def detect_arch(model_path: str, target: str = "") -> Optional[str]:
+    """Architecture key for a base model, or None when it isn't one of the new ones.
+
+    An explicit target wins; otherwise the model path/id is matched. `video` stays
+    a synonym for Wan so existing configs keep meaning what they meant.
+    """
+    t = (target or "").strip().lower()
+    if t in ARCHS:
+        return t
+    if t == "video":
+        hay = (model_path or "").lower()
+        for key, spec in ARCHS.items():
+            if key == "wan":
+                continue
+            if any(d in hay for d in spec["detect"]):
+                return key
+        return "wan"
+    hay = (model_path or "").lower()
+    for key, spec in ARCHS.items():
+        if any(d in hay for d in spec["detect"]):
+            return key
+    return None
+
+
+def load_classes(arch: str):
+    """Import the classes this architecture trains with.
+
+    Raises RuntimeError naming the missing piece — these builds differ (LTX-2 is in
+    diffusers 0.38, H3 and Krea 2 need 0.40), and a bare ImportError deep in a
+    training job tells nobody anything.
+    """
+    spec = ARCHS.get(arch)
+    if not spec:
+        raise RuntimeError(f"Unknown LoRA architecture '{arch}'")
+    import importlib
+    diffusers = importlib.import_module("diffusers")
+    missing = [n for n in (spec["transformer"], spec["vae"]) if not hasattr(diffusers, n)]
+    if missing:
+        need = spec.get("needs") or "a newer diffusers"
+        raise RuntimeError(
+            f"{spec['label']} LoRA training needs {', '.join(missing)}, which this "
+            f"diffusers ({getattr(diffusers, '__version__', '?')}) does not provide. "
+            f"Requires {need}.")
+    tr_cls = getattr(diffusers, spec["transformer"])
+    vae_cls = getattr(diffusers, spec["vae"])
+    te_mod, te_name = spec["text_encoder"]
+    tok_mod, tok_name = spec.get("tokenizer", ("transformers", "AutoTokenizer"))
+    te_cls = getattr(importlib.import_module(te_mod), te_name)
+    tok_cls = getattr(importlib.import_module(tok_mod), tok_name)
+    return tr_cls, vae_cls, te_cls, tok_cls, spec

tools/character_studio.py

@@ -1013,7 +1013,14 @@ video{width:100%;border-radius:14px;margin-top:12px;background:#000}
         <div><label>Trigger word</label><input id="trigger" placeholder="(character name)"></div>
       </div>
       <div class="row3">
-        <div><label>Target</label><select id="train_target"><option value="video">video</option><option value="image">image</option></select></div>
+        <div><label>Target</label><select id="train_target">
+          <option value="video">video (auto)</option>
+          <option value="wan">wan</option>
+          <option value="ltx2">ltx2</option>
+          <option value="h3">minimax-h3</option>
+          <option value="krea">krea 2</option>
+          <option value="image">image (sd/sdxl/z-image)</option>
+        </select></div>
         <div><label>Steps</label><input id="train_steps" type="number" value="800"></div>
         <div><label>Rank</label><input id="rank" type="number" value="16"></div>
       </div>
@@ -1454,8 +1461,12 @@ def video_opts(args: argparse.Namespace) -> dict[str, Any]:
 
 def add_train_options(parser: argparse.ArgumentParser) -> None:
     parser.add_argument("--lora-name", default="", help="Output LoRA name (default: <character>_identity)")
-    parser.add_argument("--target", choices=["video", "image"], default="video",
-                        help="Pipeline the LoRA is trained for (default: video)")
+    parser.add_argument("--target", default="video",
+                        choices=["video", "image", "wan", "ltx2", "h3", "krea"],
+                        help="Which architecture to train the LoRA for. 'video' picks "
+                             "it from the model id (Wan unless it looks like LTX-2 or "
+                             "H3); the explicit names force one. 'image' is the "
+                             "SD1.x/SDXL/Z-Image trainer. (default: video)")
     parser.add_argument("--base-model", default="", help="Model to train against (default: the studio's video/image model)")
     parser.add_argument("--trigger", default="", help="Trigger word (default: the character name)")
     parser.add_argument("--instance-prompt", default="", help="Override the training caption")