feat(ml): generación de audio en ComfyUI (ACE-Step) — comfyui_build_audio_workflow + comfyui_fetch_output_audio

python/functions/ml/tests/test_comfyui_build_audio_workflow.py

@@ -0,0 +1,90 @@
+"""Tests de estructura para comfyui_build_audio_workflow (funcion pura, ACE-Step)."""
+
+import os
+import sys
+
+sys.path.insert(0, os.path.dirname(__file__))
+sys.path.insert(0, os.path.join(os.path.dirname(__file__), "..", ".."))
+
+from ml.comfyui_build_audio_workflow import comfyui_build_audio_workflow
+from _comfyui_wf_assert import assert_api_format, class_types, node_by_ct
+
+
+def test_estructura_y_nodos_acestep():
+    wf = comfyui_build_audio_workflow(
+        "AUDIO_ace_step_v1_3.5b.safetensors", "retro coin sfx"
+    )
+    assert_api_format(wf)
+    cts = class_types(wf)
+    for ct in (
+        "CheckpointLoaderSimple",
+        "TextEncodeAceStepAudio",
+        "ConditioningZeroOut",
+        "EmptyAceStepLatentAudio",
+        "ModelSamplingSD3",
+        "KSampler",
+        "VAEDecodeAudio",
+        "SaveAudio",
+    ):
+        assert ct in cts, f"falta nodo {ct}"
+    assert len(wf) == 8
+
+
+def test_ckpt_y_prompt_reflejados():
+    wf = comfyui_build_audio_workflow("AUDIO_x.safetensors", "magic spell whoosh")
+    assert node_by_ct(wf, "CheckpointLoaderSimple")["inputs"]["ckpt_name"] == "AUDIO_x.safetensors"
+    enc = node_by_ct(wf, "TextEncodeAceStepAudio")
+    assert enc["inputs"]["tags"] == "magic spell whoosh"
+    assert enc["inputs"]["lyrics"] == ""
+
+
+def test_cableado_ksampler():
+    wf = comfyui_build_audio_workflow("AUDIO_x.safetensors", "p")
+    ks = node_by_ct(wf, "KSampler")["inputs"]
+    # model viene de ModelSamplingSD3 ("11"), no del checkpoint directo
+    assert ks["model"] == ["11", 0]
+    assert ks["positive"] == ["6", 0]
+    # negative pasa por ConditioningZeroOut ("10")
+    assert ks["negative"] == ["10", 0]
+    assert ks["latent_image"] == ["5", 0]
+    assert ks["denoise"] == 1.0
+    # ModelSamplingSD3 toma el MODEL del checkpoint
+    assert node_by_ct(wf, "ModelSamplingSD3")["inputs"]["model"] == ["4", 0]
+    # VAEDecodeAudio usa el VAE del checkpoint
+    assert node_by_ct(wf, "VAEDecodeAudio")["inputs"]["vae"] == ["4", 2]
+    # ConditioningZeroOut deriva del positive
+    assert node_by_ct(wf, "ConditioningZeroOut")["inputs"]["conditioning"] == ["6", 0]
+
+
+def test_edge_seconds_y_seed_variables():
+    wf_a = comfyui_build_audio_workflow("c", "p", seconds=4.0, seed=42)
+    wf_b = comfyui_build_audio_workflow("c", "p", seconds=8.0, seed=99)
+    assert node_by_ct(wf_a, "EmptyAceStepLatentAudio")["inputs"]["seconds"] == 4.0
+    assert node_by_ct(wf_b, "EmptyAceStepLatentAudio")["inputs"]["seconds"] == 8.0
+    assert node_by_ct(wf_a, "KSampler")["inputs"]["seed"] == 42
+    assert node_by_ct(wf_b, "KSampler")["inputs"]["seed"] == 99
+
+
+def test_params_reflejados():
+    wf = comfyui_build_audio_workflow(
+        "c", "p",
+        lyrics="la la la", steps=30, cfg=4.0, sampler_name="dpmpp_2m",
+        scheduler="karras", shift=3.5, lyrics_strength=0.7,
+        filename_prefix="audio/mio",
+    )
+    enc = node_by_ct(wf, "TextEncodeAceStepAudio")["inputs"]
+    assert enc["lyrics"] == "la la la"
+    assert enc["lyrics_strength"] == 0.7
+    ks = node_by_ct(wf, "KSampler")["inputs"]
+    assert ks["steps"] == 30
+    assert ks["cfg"] == 4.0
+    assert ks["sampler_name"] == "dpmpp_2m"
+    assert ks["scheduler"] == "karras"
+    assert node_by_ct(wf, "ModelSamplingSD3")["inputs"]["shift"] == 3.5
+    assert node_by_ct(wf, "SaveAudio")["inputs"]["filename_prefix"] == "audio/mio"
+
+
+def test_determinismo():
+    a = comfyui_build_audio_workflow("c", "p", seconds=5.0, seed=7)
+    b = comfyui_build_audio_workflow("c", "p", seconds=5.0, seed=7)
+    assert a == b

python/functions/ml/tests/test_comfyui_fetch_output_audio.py

@@ -0,0 +1,50 @@
+"""Tests de localizacion de output para comfyui_fetch_output_audio.
+
+Solo cubren la logica pura de busqueda (_is_audio_item / _find_audio_output): no
+tocan red ni disco. La descarga real via HTTP se prueba en el flujo e2e con el
+servidor ComfyUI vivo.
+"""
+import os
+import sys
+
+sys.path.insert(0, os.path.dirname(__file__))
+
+from comfyui_fetch_output_audio import _find_audio_output, _is_audio_item
+
+
+def test_is_audio_item_por_extension():
+    assert _is_audio_item({"filename": "comfy_audio_00001_.flac"})
+    assert _is_audio_item({"filename": "x.mp3"})
+    assert _is_audio_item({"filename": "x.WAV"})
+    assert not _is_audio_item({"filename": "x.png"})
+    assert not _is_audio_item({"filename": ""})
+
+
+def test_find_saveaudio_flac_bajo_audio():
+    outputs = {
+        "9": {"audio": [{"filename": "comfy_audio_00001_.flac",
+                          "subfolder": "audio", "type": "output"}]}
+    }
+    got = _find_audio_output(outputs)
+    assert got == {"filename": "comfy_audio_00001_.flac",
+                   "subfolder": "audio", "type": "output"}
+
+
+def test_find_saveaudiomp3_bajo_audio():
+    outputs = {"12": {"audio": [{"filename": "track.mp3", "subfolder": "", "type": "output"}]}}
+    assert _find_audio_output(outputs)["filename"] == "track.mp3"
+
+
+def test_find_prioriza_clave_audio():
+    # Un nodo deja un png bajo "images" y otro un flac bajo "audio": gana el audio.
+    outputs = {
+        "9": {"images": [{"filename": "preview.png", "subfolder": "", "type": "output"}]},
+        "10": {"audio": [{"filename": "out.flac", "subfolder": "", "type": "output"}]},
+    }
+    assert _find_audio_output(outputs)["filename"] == "out.flac"
+
+
+def test_find_sin_audio_devuelve_none():
+    outputs = {"9": {"images": [{"filename": "preview.png", "subfolder": "", "type": "output"}]}}
+    assert _find_audio_output(outputs) is None
+    assert _find_audio_output({}) is None