thepatch/same-l-decoder-lora

SAME-L Decoder LoRA (squeakfix_v3)

A LoRA on SAME-L's decoder. It changes how latents are rendered to audio, not how latents are produced, so the latent space, the DiT and every existing DiT LoRA stay bit-compatible and this stacks with them in any order.

It targets the high-frequency squeaks SAME's autoencoder adds to percussion when audio goes round the encode/decode loop more than once. A continuation or a transform re-encodes its input, so a chained session compounds them.

Powered by Stability AI. This is a Derivative Work of SAME-L under the Stability AI Community License; NOTICE records what it changes. No SAME-L weights are redistributed here, only the adapter.

Listen

	stock decoder	squeakfix_v3
jungle, fresh generation
jungle, after 3 round trips
drums, fresh generation
drums, after 3 round trips
neurofunk, fresh generation
neurofunk, after 3 round trips

15 s clips, same latent and seed, decoder adapter the only variable, peak-matched to −1 dBFS, generated on medium with a DiT LoRA attached. These three prompts were not picked for artifact content, and the fresh-generation rows are there to show the progression rather than because much is audible in them.

On a fresh generation

	stock decoder	squeakfix_v3
glitch hop, worst of 16 screened

Much subtler, and it took screening 16 candidates by tonality runaway to find a pair where it carries. What changes here is air rather than squeak count: on this pair tonality moves 0.08 dB, which is nothing, while 16-22 kHz comes back +2.4 dB.

How this came about

I do a lot of transformations and continuations using my guitar as input audio. When I first noticed these squeaks in the outputs, I mentioned that my trick was to stem separate in Ableton and then apply the erosion effect to the percussion layer. Claude used that as a guide for refining the methodology. The technical version is below.

Usage

Requires the target: "decoder" support on the decoder-lora branch of betweentwomidnights/stable-audio-3. It is not upstream yet, so that branch is currently the only place a decoder LoRA will load. The target field in the checkpoint config routes it; there is no special loading path. Ship at strength 1.0, where it was trained and auditioned.

python run_gradio.py --model medium-base --lora-ckpt-path squeakfix_v3.safetensors

# stacks with a DiT LoRA, either order
python run_gradio.py --model medium-base \
    --lora-ckpt-path my_style.safetensors squeakfix_v3.safetensors

from stable_audio_3 import AutoencoderModel
from stable_audio_3.models.lora.loader import load_and_apply_loras
from stable_audio_3.models.lora.model import set_lora_strength

ae = AutoencoderModel.from_pretrained("same-l", device="cuda").autoencoder
load_and_apply_loras(ae, ["squeakfix_v3.safetensors"], "autoencoder")
set_lora_strength(ae.decoder, 1.0)   # separate call; 0.0 is the stock decode

audio = ae.decode(ae.encode(waveform))

Numbers

Four base-model generations of 2 minutes on ordinary descriptive prompts, screened from 12 candidates by tonality runaway. Each latent is decoded twice, adapter off and on, and each chain re-encodes in place, so dN is N round trips of that decoder config. The adapter is the only variable.

depth	tonality p95	tonal frames	16-22 kHz	SI-SDR vs own d0
d0, fresh generation	9.80 → 9.10	87 → 62	3.16 → 4.34
d1, one continuation	11.82 → 9.73	462 → 64	0.95 → 3.76	16.9 → 21.0
d2	14.79 → 10.61	2111 → 133	−1.36 → 3.32	12.4 → 16.2
d3	17.85 → 11.74	4279 → 453	−3.36 → 2.96	10.1 → 13.6

The adapter is ahead at every depth, including a fresh generation. Stock tonality climbs +8.05 dB over three round trips against the adapter's +2.64, and by d3 stock carries 4279 tonal frames against 453.

It is not doing this by sanding the top end. After three round trips the adapter's 16-22 kHz sits 0.2 dB from a fresh stock render, at unchanged RMS, while stock has lost 6.5 dB of it. Stock's 12-16 kHz meanwhile rises with depth: the invented energy is replacing the air.

Prompt distribution matters more here than anything else. The same ladder run on terse genre-tag prompts, which are out of distribution for the base model, starts at 1649 tonal frames instead of 87. With a DiT LoRA attached, which makes those tags in-distribution again, it starts at 780. The artifact is worst where the prompt sits furthest from what the model expects, and re-encoding compounds whatever it starts with.

A DiT LoRA does not change the shape: on that run stock climbs +5.79 dB against the adapter's +0.73, and the adapter ends d3 holding more 12-16 kHz and 16-22 kHz than a fresh stock render. All runs are fp16 at 120 s.

Limits

• Decoder LoRAs trained this way stamp a comb into the output. The un-patch is a bare reshape at patch_size 256, so output channel and time-position-within-patch are the same axis, and any channel-wise bias the adapter learns repeats in every patch. That is a harmonic comb at sr/256 = 172.27 Hz, which reads as a wrong note rather than as noise. It is inaudible in a mix and obvious once a stem separator strips the masking content, which is how it was found. --lambda_patch 30 is what prevents it: a v1 control trained without it sits +10.9 dB of comb excess over stock while v3 sits 0.8 dB below stock, and both hold through three further round trips. On real audio rather than model output v3 measures +1.5 to +4.0 dB, so measure on generated audio.
• SAME-L only. SAME-S shares the patch grid but its artifact sits 1-8 kHz and wants a different recipe. Loading this there will not error, it will sound wrong.
• The gain scales with re-encode depth, so it is small for one-shot generation and compounds across chained continuations.
• The squeaks are worst on out-of-distribution prompts. Genre-tag prompts on the base model start with 19x more tonal frames than descriptive ones. A DiT LoRA makes such tags in-distribution again and lands in between. This adapter reduces the artifact; it does not make prompt choice irrelevant.
• One checkpoint, one dataset, step 2000, chosen on re-encode behaviour and listening. Your material may want a different one.
• Neither half of the autoencoder is bit-deterministic (encoder 1.5 % relative, decoder 7e-3 max abs), so read any A/B as a multiple of those floors.

Training

Self-supervised, audio only, no captions or paired data. Rank 16 / alpha 16, 5.63M trainable params (~1.30 % of the decoder), fp32 training and fp16 storage, single GB10 (DGX Spark) at ~2.9 s/step on 10 s crops, under 10 GB.

python scripts/train_decoder_lora.py \
    --model same-l --rank 16 --steps 8000 --lr 1e-4 \
    --w_real 0.4 --w_drift 0.3 --w_dit 0.3 \
    --lambda_cycle 10.0 --lambda_tonal 0.3 --lambda_patch 30 \
    --patch_size 256 --crop_seconds 10 --seed 0 \
    --dit_latent_dir out/dit_latents \
    --data_dir /path/to/audio --eval_audio /path/to/holdout.wav \
    --out_dir out/declora

Latents come from three buckets: real audio, drifted audio (E(D(E(x))) at depth ≥ 1, targeted against the original so the decoder is asked to undo accumulated drift), and latents sampled from the base DiT. No discriminator, since adversarial training makes a decoder invent more detail.

Full method, losses, ablations and traps: Decoder LoRA workflow. Provenance is embedded in the file's lora_config metadata. v1 (identical recipe, --lambda_patch 0) is an internal control, not published.

Trained and used by the collabage patch, in production in a live stable-audio-3 backend since 2026-08-14. The autoencoder this adapts: SAME.