LoRA & TTS Studio · Model adaptation workstation

Fine signature + manual review. Problem: High scores in multi-dimensional automatic scoring are still mixed with covered eyes/profiles/faces, and the original labels of the data set are also dirty. Solution: Automatic scoring → top candidates → manual review to keep clean exemplars, scalar vectors as true values, and discard the original labels of the data set.

Multi-dimensional data screening. Problem: Single indicator (sharpness) cannot capture composition/expression/diversity in film selection, and 80% of high scores are false positives. Solution: Segmentation model to calculate occlusion + hash deduplication + angle soft gate + perceptual deduplication, real-time dynamic weighting, low memory machine limit concurrency explosion prevention.

The pod basic environment is unified. Problem: The bare image lacks common libraries, and multiple bootstrap scripts drift separately. Solution: The unified baseline script automatically completes dependencies, and the channel dynamically pulls versions, which is compatible with multiple generations of pods.

TTS training environment isolation. Problem: The deep learning library version required by the TTS framework conflicts with image training, and the port is occupied by the default service. Solution: The independent virtual environment is not shared with the main environment, port offset is used to avoid conflicts, and the entire environment is installed on the local disk.

LoRA dim dimension empirical fallback. Problem: The textbook route believes that the higher the dimension, the more details can be contained, so adjust it to dim 64. Solution: Use the same seed and prompt to evaluate 10 pictures, dim 32 full is inversely proportional to dim 64, which is closer to me, and the high capacity can learn noise in small data sets. Rewind dim ≤ 32, and the control variable is evaluated as a rule.

Automatic image verification panel after training. Problem: After training, you can only manually pull the weights to run the local image for verification, and the feedback link is too long. Solution: Add an API endpoint to directly drive the remote image verification, use the transit channel to export the image, and the front-end skeleton loading experience.

Hugging Face relay replaces SFTP direct transmission. Problem: The single thread of the original direct transmission protocol crushes the bandwidth (14 times slower), the marking tool is loaded into the main environment and pollutes the deep learning library, and the root disk 17 GB is full of logs. Solution: Data/models/logs are uniformly transferred through Hugging Face, the marking tool isolates the virtual environment, and the cache directory is diverted to avoid root disk explosion.

The training and graph verification co-locate in the same pod. Problem: The deep learning library versions of the training framework and the graph verification framework conflict, and GPU memory contention. Solution: Isolate the virtual environment with independent site packages, model paths point to the same directory, and multi-layer smoke passes.

Multi-base model hyperparameter profile. Problem: The optimal learning rate, noise, and SNR coefficient of different base models (SDXL / Flux / RealVisXL, etc.) vary greatly, and manual switching is prone to errors. Solution: The built-in realistic mode switch links a set of hyperparameters with one click, and the real machine's video memory usage is stable and does not overflow.