Spaces:
Running on Zero
Running on Zero
Commit History
-
Upload lora_loader.py with huggingface_hub
Verified signaturec3d3fe5
-
Smashcut: drop ambiguous 'in-and-out' require (false-positives on fingering)
The bare phrase 'in-and-out' normalized to 'in and out' and matched finger-thrusting scenes (e.g. 'His fingers slide in and out'), causing Smashcut (camera) to inject penis-penetration content into a hands-only fingering scene. Removed it; all doggystyle/bench/machine scenarios still match Smashcut via the remaining specific requires.8f58600 -
Fix cross-LoRA trigger token reuse (BBC-class audit)
Audit of LORA_PROMPT_TRIGGERS against official triggerword docs (rolechar .txt files, lopi999 card, Civitai) found two more LoRAs prepending another LoRA's activation token: - Doggy Front View v2 was prepending iGoon's 'front doggystyle' token; documented trigger is 'Pl0wView'. - Fucked From Behind was prepending sfbehind's 'sex from behind' phrase; its own token is 'fucked from behind' (matches derpfuckedfrombehind filename and reference prompt). All other triggers verified already-correct (facial, sensualBJ, nsfwsks, m0u7hfu11, sensual_fingering, pornmaster bukkake, iGoon doggystyle tokens, FOV slider camera phrase). JFJ Deepthroat stays on 'blowjob, deepthroat' (its published trigger).7fe59f2 -
BBC Deepthroat: use the bbc deepthroat trigger token
The prepended trigger was 'deep throat blowjob' — the generic Ultimate Deepthroat token — so the BBC-specific trigger never reached the prompt even though the BBC LoRA had fired. The trained trigger (per the reference prompt and phrase list) is 'bbc deepthroat'; the combined prompt now opens with 'bbc deepthroat; nsfwsks, ...'. BBC Blowjob Extreme already used its own 'bbc extreme blowjob' token.563c2df -
BBC LoRAs: fire on black-context oral scenes
'black man' previously matched no trigger phrase, so a black-man deepthroat/facefuck scene fell through to the generic Ultimate Deepthroat LoRA instead of the BBC variant. BBC Deepthroat and BBC Blowjob Extreme now share the generic action phrases of their category (supersets), but are gated via LORA_AUTO_REQUIRES on black context (black man/black cock/bbc/ interracial). Because the BBC entries sit after the generic ones in the dict, they win the oral-category tie on black scenes, while the gate keeps them excluded from non-black scenes (plain deepthroat, plain blowjob, kissing scenes mentioning a black man all fall through to the generic LoRA).40bf80e -
Deepthroat: add verb-form facefuck phrases
The blowjob LoRA additions were winning the winner-take-all 'oral' category and replacing Ultimate Deepthroat on full-throat scenes, regressing log behavior (BJ deepthroat scenes fire Booster+Deepthroat). Move the aggressive facefuck verb phrases (fucks her face/mouth, fucking her face/mouth, cock fully into her mouth, fully into her mouth, deep in her mouth) onto Ultimate Deepthroat, where they belong: the dedicated blowjob LoRA stays reserved for generic sucking scenes.6ef364a
-
Consistent Smashcut firing on hard-cut thrusting scenes; fix breast-grope activation
Audit of the production log (07b1e48) exposed two activation gaps: 1. Smashcut fired on only 2 of 6 nearly-identical doggystyle hard-cut scenes. The requires gate matched exact phrase forms only: 'fucking her' fired but 'fucks her quickly' didn't; 'in and out of her pussy' fired but 'fully in and fully out'/'in-and-out strokes' didn't. Added verb-form variants ('fucks her pussy', 'thrusting hard', 'thrusting fast', 'fully in and fully out', 'in-and-out', 'driving his hips forward', 'hips driving forward', 'hips slamming forward'). All 6 doggystyle variants + the bench-penetration scene now fire Smashcut consistently. Mouth-fucking scenes stay clean (no bare 'fucks her'). 2. Pinata/grope scenes ('caresses her breasts... groping, squeezing, and stroking them') fired only the generic booster. Added massage-tits phrase variants ('caresses her breasts', 'groping her breasts', 'groping them', 'groping and squeezing', 'stroking her breasts', 'stroking the…f3a961d -
Fix LoRA load failure for Wan-native diff_b format files
Production log: 'iGoon - Blink_Front_Doggystyle_I2V' and 'Ultimate Deepthroat' failed to load with 'TorchaoLoraLinear does not support lora_bias yet', and generation proceeded WITHOUT their weights while the prompt tags still claimed they were active. Root cause: these LoRAs are saved in the Wan-native diff_b format (447 'diffusion_model.*.diff_b' learned-bias keys per file). diffusers' _convert_non_diffusers_wan_lora_to_diffusers maps each diff_b key to 'lora_B.bias', which flips lora_bias=True, which peft's TorchaoLoraLinear hard-rejects on torchao-quantized layers. The previous strip only removed literal 'lora_B...bias' keys from already-converted dicts, missing the Wan-native form entirely. Fix: _strip_lora_bias now also drops keys ending in '.diff_b'. Verified against the real production files (447/447 bias terms stripped, weights preserved) and added a mock-pipe test case for the Wan-native format. Regression: 5/5 mock-pipe cases, 18/18 auto-activation matrix, py_compile all p…07b1e48 -
Auto-LoRA: fire front doggystyle + Smashcut on hard-cut doggy scene
The surveillance doggystyle prompt (on all fours, facing the camera, roughly fucking her) only activated General NSFW Booster. Added: - iGoon - Blink_Front_Doggystyle_I2V: 'doggystyle position', 'doggystyle rhythm', 'facing toward the camera', 'facing towards the camera' (the prompt's phrasing was 'on all fours in a doggystyle position, facing toward the camera' — no directional variant previously matched) - Smashcut requires gate: 'fucking her' so a hard cut into a fucking scene passes the thrusting-scene gate Verified: prompt now activates Front Doggystyle + Smashcut + General NSFW Booster; full 18-test matrix, previous penetration/creampie prompt, and LoRA-load tests all still pass with no new false positives.429d0f4 -
Extend auto-LoRA coverage for penetration/thrusting/creampie phrasing
The surveillance-basement prompt (hard cut + forced penetration + creampie) only activated General NSFW Booster. Added phrases for the phrasings real prompts use: - Penetration Insert: 'into her pussy', 'thrusting in and out', 'pounding her hard', 'pounding her pussy', 'pounds her pussy' - Smashcut requires gate: 'thrusting in and out', 'pounding her hard' so a hard cut followed by thrusting/pounding fires it - Creampie CRM: 'finishes deep inside her', 'finish deep inside her', 'finishing inside her', 'leaking out of her', 'cum leaking out', 'cum visible leaking' Verified: the reported prompt now activates Penetration Insert + Smashcut + Creampie CRM + General NSFW Booster; 18-test regression matrix still passes with no new false positives (incl. 'no zooming' negation and identity boilerplate cases).15712cb
-
Fix runtime LoRA loading on torchao-quantized pipe (bias + activation + cleanup)
Three bugs seen in production logs after the peft 0.20.0 upgrade: 1. 'TorchaoLoraLinear does not support lora_bias yet, set it to False' — diffusers auto-detects lora_bias=True purely from lora_B.*.bias keys in the LoRA state dict (peft_utils.py), and peft's TorchaoLoraLinear hard-rejects it on torchao-quantized layers. load_lora_to_pipe now loads via the dict path (safetensors.torch.load_file) and strips lora_B bias keys first, so such LoRAs (e.g. Ultimate Deepthroat) load successfully. 2. Adapter-name collision cascade ('Adapter name lora_0_high already in use') — a partial failure left lora_loaded=False so unload_lora was skipped and stale adapters leaked into the next task. load_lora_to_pipe now unloads all adapters on failure and re-raises; app.py also unloads in the except path, keeping adapter names reusable across tasks. 3. Latent multi-LoRA activation bug — each load_lora_weights leaves only the last-loaded adapter active, so multiple auto-acti…bf209fb -
Fix auto-LoRA prompt matching: negation guard, typo fixes, phrase coverage
- Add _is_negated guard so negated camera/zoom phrases ('no zooming', 'no zooms that crop') don't fire FOV, while positive 'zoom-ins' still do - Add _TYPO_FIXES map in _normalize (naaked, nuude, niipples, vaagina, ...) - FOV: drop bare 'zoom', require zoom in/out/zooms in/zooms out - Remove over-broad phrases: 'spreading her' (Pussy Helper), bare 'facial' (iGoon Blink_Facial), 'standing position' (Standing Upright Sex), 'bent over' (sfbehind) - Expand coverage: breast squeeze/cup/caress + pronoun forms, fingering masturbation phrases, deep penetration wording, vulva/labia close-up, nudity phrases mapped to General NSFW Booster - Smashcut requires: add 'into and out of her pussy' so 'thrusts into and out of her pussy' phrasing gates it correctly - Tie-break: last-scored group wins on exact score tie (iGoon Facial over Facesplash on shared 'facial cumshot')57207b5 -
Rebuild auto LoRA activation: ICF-scored prompt matching + live preview
New architecture (not a restore of the prior implementation): - Normalized matching (hyphen/space-insensitive, punctuation-stripped) so 'face-fucking' == 'face fucking' - Inverse-concept-frequency scoring: distinctive phrases (unique to few LoRAs) score high, generic phrases shared across many score low; replaces the old longest-keyword-wins heuristic with a principled signal - Category dedupe by highest score (one LoRA per concept), Smashcut gated on pussy-thrusting scene via LORA_AUTO_REQUIRES - auto_activation_details() returns matched phrases + score for a live, explainable preview panel under the prompt box (checkbox default ON) - Data-completeness: generic sex prompts fire the NSFW Booster; thrusting into pussy fires Penetration Insert - Prompt adaptation ported: trigger-words-only merge (user scene wins), full body kept for camera/booster LoRAsbb9df7f
-
copy of thornmaze/wan2-2-i2v-v3 (byte-1:1)
Verified signature1b93f28