Jason HokuandClaude Opus 4.7
4fb4fea5ed
When the user types '0.5, 1.0' they want 2 outputs (lora at 0.5, lora at 1.0 — both model and clip mirroring), not 4 outputs (Cartesian m×c). The bug was that we wrote both _model AND _clip side-channel arrays, so Python emitted 'name:[m]:[c]' which the orchestrator correctly interpreted as Cartesian. JS: collapse the two-input compare section to a single 'Strengths:' input. Never write _clip to lora_weight_arrays. Python: process_lora_array now emits 'name:[m_arr]' (no clip part) when only _model array is present. The generator's existing parse_lora_definition defaults clip=model when the third segment is missing, giving N configs with locked semantics. Drops the Cartesian path entirely. Backward compat retained: if a saved workflow has both arrays (legacy), the dual-bracket form is still emitted so the orchestrator can Cartesian-expand as before. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
203 lines
8.5 KiB
Python
203 lines
8.5 KiB
Python
"""Golden tests for UltimateConfigBuilder.state_to_configs_json.
|
||
|
||
This is the single transformer that both the runtime (generate_config) and
|
||
the preview endpoint (/configbuilder/preview) call. Drift between preview
|
||
and runtime is impossible by construction because both call this function.
|
||
"""
|
||
import json
|
||
from config_builder_node import UltimateConfigBuilder
|
||
|
||
|
||
def make_state(loras=None, lora_weight_arrays=None, lora_bypass_states=None,
|
||
models=None, samplers=None, schedulers=None, steps="20", cfg="7.0",
|
||
extra_array_fields=None):
|
||
"""Build a minimal valid state with one config_array for tests."""
|
||
arr = {
|
||
"name": "Test",
|
||
"samplers": samplers or ["euler"],
|
||
"schedulers": schedulers or ["normal"],
|
||
"steps": steps,
|
||
"cfg": cfg,
|
||
"models": models or ["None"],
|
||
"vaes": ["None"],
|
||
"text_encoders": [],
|
||
"clip_type": "stable_diffusion",
|
||
"gguf_options": {},
|
||
"loras": loras if loras is not None else ["None"],
|
||
"lora_omit_triggers": [],
|
||
"lora_triggerwords_append_settings": {},
|
||
"lora_bypass_states": lora_bypass_states or {},
|
||
"lora_strength_lock": {},
|
||
"lora_weight_arrays": lora_weight_arrays or {},
|
||
"model_bypass_states": {},
|
||
"vae_bypass_states": {},
|
||
"te_bypass_states": {},
|
||
"combine": False,
|
||
"positive_prompt_groups": [],
|
||
"negative_prompt": "",
|
||
"use_custom_prompts": False,
|
||
"model_prompt_prefix": "",
|
||
"model_prompt_suffix": "",
|
||
"attention_modes": ["default"],
|
||
}
|
||
if extra_array_fields:
|
||
arr.update(extra_array_fields)
|
||
return {
|
||
"session_name": "test_session",
|
||
"include_none": False,
|
||
"global_positive_groups": [],
|
||
"global_negative": "",
|
||
"config_arrays": [arr],
|
||
}
|
||
|
||
|
||
def test_lora_strength_arrays_render_as_brackets():
|
||
"""REGRESSION: Compare Strengths emits single-bracket form for locked semantics.
|
||
|
||
Bug history:
|
||
- 2026-04-28 (morning): JS preview emitted brackets but Python emitted
|
||
scalar 1.00:1.00 because Python ignored lora_weight_arrays side-channel.
|
||
- 2026-04-28 (afternoon): Both sides emitted "name:[m]:[c]" (two arrays),
|
||
which the orchestrator interpreted as Cartesian N×M configs instead of
|
||
N. Locked-mode fix: emit only "name:[m]" (no clip part) so the orchestrator
|
||
expands to N configs each with model=clip=value.
|
||
|
||
State setup uses ONLY lora_weight_arrays[name + "_model"] (no _clip), which
|
||
is the data shape the Builder UI now writes for compare strengths.
|
||
"""
|
||
state = make_state(
|
||
loras=["Sexy-IL-v11/:1.00:1.00"],
|
||
lora_weight_arrays={
|
||
"Sexy-IL-v11/_model": [0, 1, 2, 5, 10],
|
||
},
|
||
)
|
||
json_str = UltimateConfigBuilder.state_to_configs_json(state)
|
||
parsed = json.loads(json_str)
|
||
loras = [c["lora"] for c in parsed["configs"]]
|
||
# Locked semantics (default): single bracket array, no clip part.
|
||
# The orchestrator expands "name:[a,b,c]" to N configs each with
|
||
# model=clip=value (parse_lora_definition defaults clip to model
|
||
# when the third segment is missing).
|
||
expected = "Sexy-IL-v11/:[0, 1, 2, 5, 10]"
|
||
assert any(l == expected for l in loras), \
|
||
f"Expected exact lora string {expected!r} in output. Got: {loras}"
|
||
|
||
|
||
def test_single_lora_passes_through_verbatim():
|
||
"""A single LoRA with scalar strengths shows up unchanged in configs_json."""
|
||
state = make_state(loras=["myLora:0.80:0.50"])
|
||
json_str = UltimateConfigBuilder.state_to_configs_json(state)
|
||
parsed = json.loads(json_str)
|
||
assert len(parsed["configs"]) == 1, parsed["configs"]
|
||
assert parsed["configs"][0]["lora"] == "myLora:0.80:0.50"
|
||
|
||
|
||
def test_multiple_loras_combine_with_plus_separator():
|
||
"""combine: True (forced) joins multiple stackable loras with ' + '.
|
||
|
||
Note: process_lora_array forces combine=True regardless of the field's
|
||
value in the config_array (see config_builder_node.py 'combine = True').
|
||
"""
|
||
state = make_state(loras=["lora1:0.5:1.00", "lora2:0.7:1.00"])
|
||
json_str = UltimateConfigBuilder.state_to_configs_json(state)
|
||
parsed = json.loads(json_str)
|
||
assert parsed["configs"][0]["lora"] == "lora1:0.5:1.00 + lora2:0.7:1.00"
|
||
|
||
|
||
def test_folder_reference_left_unexpanded():
|
||
"""Folder refs ('myFolder/') stay literal — orchestrator expands at runtime."""
|
||
state = make_state(loras=["myFolder/"])
|
||
json_str = UltimateConfigBuilder.state_to_configs_json(state)
|
||
parsed = json.loads(json_str)
|
||
lora = parsed["configs"][0]["lora"]
|
||
assert lora == "myFolder/", f"Folder ref should not be expanded here, got: {lora}"
|
||
|
||
|
||
def test_folder_with_strength_array_not_combined_with_other_loras():
|
||
"""A folder ref with strength arrays must keep its trailing '/' AND
|
||
bracket form, AND must NOT be combined with stackable loras (the combine
|
||
block uses 'name part ends with /' to detect folder refs)."""
|
||
state = make_state(
|
||
loras=["myFolder/", "actualLora:0.5:1.00"],
|
||
lora_weight_arrays={
|
||
"myFolder/_model": [0.5, 0.8, 1.0],
|
||
},
|
||
)
|
||
json_str = UltimateConfigBuilder.state_to_configs_json(state)
|
||
parsed = json.loads(json_str)
|
||
# Multiple loras (one folder, one stackable, no stacking happens
|
||
# because there's only one stackable), so the per-config "lora" field
|
||
# is a LIST — not a combined string.
|
||
lora = parsed["configs"][0]["lora"]
|
||
assert isinstance(lora, list), f"Expected list of separate loras, got: {lora!r}"
|
||
folder_entry = [s for s in lora if s.startswith("myFolder/")]
|
||
assert folder_entry, f"Folder ref missing: {lora}"
|
||
assert "[0.5, 0.8, 1.0]" in folder_entry[0], f"Bracket form missing: {folder_entry[0]}"
|
||
assert "actualLora:0.5:1.00" in lora, f"Stackable lora missing: {lora}"
|
||
|
||
|
||
def test_bypassed_lora_is_excluded():
|
||
"""A LoRA listed in lora_bypass_states with True is filtered out."""
|
||
state = make_state(
|
||
loras=["lora1:0.5:1.00", "lora2:0.7:1.00"],
|
||
lora_bypass_states={"lora1": True},
|
||
)
|
||
json_str = UltimateConfigBuilder.state_to_configs_json(state)
|
||
parsed = json.loads(json_str)
|
||
lora = parsed["configs"][0]["lora"]
|
||
# Only lora2 survives the bypass filter.
|
||
assert lora == "lora2:0.7:1.00", f"Expected only lora2 to survive bypass, got: {lora}"
|
||
|
||
|
||
def test_ltx_video_config_model_type_propagates():
|
||
"""An LTX video config_array (model with type='ltx_video') reaches
|
||
configs_json with model_type=='ltx_video'.
|
||
See project memory: model_type lives PER-MODEL inside config_array.models[i].
|
||
"""
|
||
state = make_state(
|
||
models=[{"path": "ltx-2.3.safetensors", "type": "ltx_video"}],
|
||
extra_array_fields={
|
||
"ltx_video": {
|
||
"clip_models": ["gemma_3_12B.safetensors", "ltx-2.3_text_projection.safetensors"],
|
||
"vae_video": "ltx_video_vae.safetensors",
|
||
"vae_audio": "ltx_audio_vae.safetensors",
|
||
"latent_upscaler": "ltx_latent_upscaler.safetensors",
|
||
"duration_seconds": 4,
|
||
"frame_rate": 24,
|
||
"sampler_stage1": "euler",
|
||
"sigmas_stage1": "1.0,0.5,0.0",
|
||
"image_strength_stage1": 0.0,
|
||
"sampler_stage2": "euler",
|
||
"sigmas_stage2": "0.5,0.25,0.0",
|
||
"image_strength_stage2": 0.0,
|
||
"img_compression": 75,
|
||
"input_image": "",
|
||
"audio_mode": "on",
|
||
},
|
||
},
|
||
)
|
||
json_str = UltimateConfigBuilder.state_to_configs_json(state)
|
||
parsed = json.loads(json_str)
|
||
cfg = parsed["configs"][0]
|
||
assert cfg["model_type"] == "ltx_video", f"Expected model_type=ltx_video, got: {cfg.get('model_type')!r}"
|
||
|
||
|
||
def test_legacy_dual_array_preserves_cartesian_form():
|
||
"""Backward compat: if a workflow was saved with separate _model and _clip
|
||
arrays (pre-2026-04-28-pm format), the bracket-fold should still emit
|
||
"name:[m]:[c]" so the orchestrator can Cartesian-expand it. New UI never
|
||
writes _clip, but legacy state should still round-trip correctly."""
|
||
state = make_state(
|
||
loras=["legacyLora:1:1"],
|
||
lora_weight_arrays={
|
||
"legacyLora_model": [0.5, 1.0],
|
||
"legacyLora_clip": [0.7, 1.0],
|
||
},
|
||
)
|
||
json_str = UltimateConfigBuilder.state_to_configs_json(state)
|
||
parsed = json.loads(json_str)
|
||
loras = [c["lora"] for c in parsed["configs"]]
|
||
expected = "legacyLora:[0.5, 1.0]:[0.7, 1.0]"
|
||
assert any(l == expected for l in loras), \
|
||
f"Expected legacy dual-array form {expected!r}. Got: {loras}"
|