feat(ml): núcleo subsistema comfyui-skill + ask_llm_vision

Grupo nuevo comfyui-skill: recetas versionadas de generación ComfyUI que compilan a un workflow cambiando solo el subject. - comfyui_build_skill_workflow (pura): receta -> workflow API format, despacha base (txt2img/flux/sdxl_refiner), sustituye {subject}+triggers, encadena loras e inject blocks (facedetailer, hires_fix). SkillWorkflowError tipada. - comfyui_inject_hires_fix (pura): inyecta 2ª pasada UltimateSDUpscale sobre dict. - comfyui_save/load/list_skill (impuras): CRUD de la librería en disco con versionado por snapshots, round-trip idéntico, filtro NSFW. - ask_llm_vision (core, claude-direct): pregunta multimodal imagen+texto via API directa Anthropic, para puntuar generaciones. - Página madre docs/capabilities/comfyui-skill.md con schema canónico de recipe.json. Tests offline: 11 verdes (6 builder + 5 inject_hires_fix). Sin GPU.
2026-06-24 14:35:46 +02:00
parent e8a66f0dad
commit 70d541fca9
15 changed files with 1666 additions and 0 deletions
@@ -0,0 +1,102 @@
+"""Tests de estructura para comfyui_build_skill_workflow (funcion pura).
+
+Compila recetas de skill a workflows ComfyUI en API format y verifica que:
+- el golden (txt2img + 1 LoRA + facedetailer) produce un dict bien formado con los
+  class_types esperados y el subject sustituido,
+- el edge (sin loras ni blocks) produce el workflow base minimo,
+- los params (seed/steps/cfg) se reflejan,
+- los error paths lanzan SkillWorkflowError (base desconocido / base que pide imagen).
+Offline: no toca GPU ni server.
+"""
+
+import os
+import sys
+
+sys.path.insert(0, os.path.dirname(__file__))
+sys.path.insert(0, os.path.join(os.path.dirname(__file__), "..", ".."))
+
+import pytest  # noqa: E402
+
+from ml.comfyui_build_skill_workflow import build_skill_workflow, SkillWorkflowError  # noqa: E402
+from _comfyui_wf_assert import assert_api_format, class_types, node_by_ct  # noqa: E402
+
+
+def _recipe_base(**over):
+    r = {
+        "schema_version": 1,
+        "slug": "portrait_cinematic_sdxl",
+        "version": "1.0.0",
+        "base_workflow": "txt2img",
+        "checkpoint": "juggernaut_xl_v11.safetensors",
+        "loras": [],
+        "params": {"steps": 30, "cfg": 5.5, "sampler_name": "dpmpp_2m",
+                   "scheduler": "karras", "width": 832, "height": 1216},
+        "prompt_scaffold": {"positive": "cinematic portrait of {subject}, sharp focus",
+                            "negative": "blurry, lowres", "trigger_words": []},
+        "blocks": [],
+        "provenance": {"source": "manual", "nsfw": False},
+    }
+    r.update(over)
+    return r
+
+
+def test_golden_txt2img_lora_facedetailer():
+    recipe = _recipe_base(
+        loras=[{"name": "add_detail.safetensors", "strength_model": 0.6, "strength_clip": 0.6}],
+        blocks=[{"type": "facedetailer", "params": {"denoise": 0.45}}],
+    )
+    wf = build_skill_workflow(recipe, "a woman with red hair", seed=42)
+    assert_api_format(wf)
+    cts = class_types(wf)
+    assert "CheckpointLoaderSimple" in cts
+    assert "KSampler" in cts
+    assert "LoraLoader" in cts          # el LoRA se inyecto
+    assert "FaceDetailer" in cts        # el bloque facedetailer se aplico
+    assert "SaveImage" in cts
+    # seed propagada al KSampler base.
+    assert node_by_ct(wf, "KSampler")["inputs"]["seed"] == 42
+    # subject sustituido en algun CLIPTextEncode positivo.
+    textos = [n["inputs"]["text"] for n in wf.values() if n["class_type"] == "CLIPTextEncode"]
+    assert any("a woman with red hair" in t for t in textos)
+    # el LoRA respeta la fuerza de la receta.
+    lora = node_by_ct(wf, "LoraLoader")["inputs"]
+    assert lora["lora_name"] == "add_detail.safetensors"
+    assert lora["strength_model"] == 0.6
+
+
+def test_edge_sin_loras_ni_blocks_da_workflow_base_minimo():
+    wf = build_skill_workflow(_recipe_base(), "a red apple", seed=7)
+    assert_api_format(wf)
+    assert class_types(wf) == {
+        "CheckpointLoaderSimple", "CLIPTextEncode", "EmptyLatentImage",
+        "KSampler", "VAEDecode", "SaveImage",
+    }
+    assert "LoraLoader" not in class_types(wf)
+    assert "FaceDetailer" not in class_types(wf)
+
+
+def test_params_y_trigger_words_se_reflejan():
+    recipe = _recipe_base()
+    recipe["prompt_scaffold"]["trigger_words"] = ["masterpiece", "8k"]
+    wf = build_skill_workflow(recipe, "a cat", seed=3)
+    ks = node_by_ct(wf, "KSampler")["inputs"]
+    assert ks["steps"] == 30 and ks["cfg"] == 5.5 and ks["seed"] == 3
+    lat = node_by_ct(wf, "EmptyLatentImage")["inputs"]
+    assert lat["width"] == 832 and lat["height"] == 1216
+    pos = [n["inputs"]["text"] for n in wf.values() if n["class_type"] == "CLIPTextEncode"]
+    assert any(t.startswith("masterpiece, 8k") for t in pos)
+
+
+def test_error_base_workflow_desconocido():
+    with pytest.raises(SkillWorkflowError):
+        build_skill_workflow(_recipe_base(base_workflow="diffusion_xyz"), "x")
+
+
+def test_error_base_que_requiere_imagen():
+    with pytest.raises(SkillWorkflowError):
+        build_skill_workflow(_recipe_base(base_workflow="img2img"), "x")
+
+
+def test_error_recipe_no_dict():
+    with pytest.raises(SkillWorkflowError):
+        build_skill_workflow(["no", "dict"], "x")
@@ -0,0 +1,71 @@
+"""Tests de estructura y pureza para comfyui_inject_hires_fix (funcion pura)."""
+
+import os
+import sys
+
+import pytest
+
+sys.path.insert(0, os.path.dirname(__file__))
+sys.path.insert(0, os.path.join(os.path.dirname(__file__), "..", ".."))
+
+from ml.comfyui_build_txt2img_workflow import comfyui_build_txt2img_workflow
+from ml.comfyui_inject_hires_fix import comfyui_inject_hires_fix
+from _comfyui_wf_assert import assert_api_format, class_types
+
+
+def _node_id(wf, ct):
+    return next(nid for nid, n in wf.items() if n["class_type"] == ct)
+
+
+def test_no_muta_la_entrada():
+    base = comfyui_build_txt2img_workflow("ck.safetensors", "POS", "NEG")
+    base_antes = {k: dict(v) for k, v in base.items()}
+    _ = comfyui_inject_hires_fix(base)
+    # La copia profunda garantiza que el dict original queda intacto (pureza).
+    assert "UltimateSDUpscale" not in class_types(base)
+    assert "UpscaleModelLoader" not in class_types(base)
+    assert set(base) == set(base_antes)
+
+
+def test_inserta_ultimatesdupscale_y_loader():
+    base = comfyui_build_txt2img_workflow("ck.safetensors", "POS", "NEG")
+    inj = comfyui_inject_hires_fix(base)
+    assert_api_format(inj)
+    assert "UltimateSDUpscale" in class_types(inj)
+    assert "UpscaleModelLoader" in class_types(inj)
+
+
+def test_repunta_el_saveimage_al_upscale():
+    base = comfyui_build_txt2img_workflow("ck.safetensors", "POS", "NEG")
+    inj = comfyui_inject_hires_fix(base)
+    upscale_id = _node_id(inj, "UltimateSDUpscale")
+    save = next(n for n in inj.values() if n["class_type"] == "SaveImage")
+    # El SaveImage debe tomar la imagen del UltimateSDUpscale, no ya del VAEDecode.
+    assert save["inputs"]["images"][0] == upscale_id
+
+
+def test_params_reflejados():
+    base = comfyui_build_txt2img_workflow("ck.safetensors", "POS", "NEG")
+    inj = comfyui_inject_hires_fix(
+        base, upscale_by=2.5, denoise=0.33, seed=123, upscale_model="x4.pth"
+    )
+    upscale_id = _node_id(inj, "UltimateSDUpscale")
+    loader_id = _node_id(inj, "UpscaleModelLoader")
+    up_in = inj[upscale_id]["inputs"]
+    assert up_in["upscale_by"] == 2.5
+    assert up_in["denoise"] == 0.33
+    assert up_in["seed"] == 123
+    assert inj[loader_id]["inputs"]["model_name"] == "x4.pth"
+    # Defaults fijos copiados del builder hermano.
+    assert up_in["mode_type"] == "Linear"
+    assert up_in["force_uniform_tiles"] is True
+    assert up_in["tiled_decode"] is False
+
+
+def test_lanza_valueerror_sin_vaedecode():
+    # Workflow sin VAEDecode: la funcion no puede localizar la fuente de imagen.
+    bad = {
+        "4": {"class_type": "CheckpointLoaderSimple", "inputs": {"ckpt_name": "x"}},
+    }
+    with pytest.raises(ValueError):
+        comfyui_inject_hires_fix(bad)