TESSRAL Enables Custom Image Model Training on M2 MacBook Air Hardware
Laptop users can now create custom models without renting A100s because TESSRAL supports local execution on an M2 MacBook Air with 16 GB RAM. Running pip install tessral allows for training without relying on Colab or cloud services. This efficiency comes from freezing the backbone and updating a tiny adapter layer, which reduces trainable parameters to 40 MB. A 50-image dataset completes in under 20 minutes on this hardware. Voice cloning uses a similar method, where 15 minutes of training on 3 minutes of clean audio produces a usable LoRA for TTS pipelines.
Peak memory on Apple Silicon remains under 10 GB without VRAM spikes. The system only needs 2–5 minutes of audio or 30–50 images, avoiding the thousands of samples required for full fine-tuning. Exported models use config JSON and standard safetensors, ensuring compatibility with Automatic1111, ComfyUI, or other LoRA-accepting engines. Users can publish a dataset hash to let others train locally without uploading raw files.
The tool currently lacks built-in evaluation, meaning users must manually test exports instead of using CLIP or FID scores. Non-English accents degrade after 10 minutes of inference due to limited phoneme coverage, though v0.4 will introduce a multilingual backbone. The interface is CLI-based, and the Gradio UI remains hidden behind a flag.
For those seeking commercial tools, VEO 3 is 20 % off, Seedance is 20 % off, and Seedream is 10 % off through October 8. WAN is 10 % off through October 4, with discounts applying automatically to every plan. These Limited-time deals accompany the release of Sensei, an AI creative director that plans shots and costs, maintains visual consistency across faces and lighting, and filters output to show only passing results.
All Replies （3）
Want a live back-and-forth? Join the global AI chat room — login to talk.
Managed to train a LoRA on my M2 Air overnight. Did it take you guys this long to finish? I was skeptical at first too — every other project either needed a 3090 or gave toy results, but this one actually runs on my 16 GB M2 MacBook Air with just pip install tessral. The trick is freezing the backbone and training only a tiny adapter layer (~40 MB of parameters instead of gigabytes), so a 50-image dataset finishes in under 20 minutes and 3 minutes of clean audio gives you a usable voice LoRA in 15 minutes. Peak memory stays under 10 GB on Apple Silicon, and the models export as standard safetensors + config JSON that drop straight into ComfyUI or Automatic1111. What surprised me most is that even on my Air, it doesn't OOM like the others did.
Starting your first fine-tune can be overwhelming, especially with VRAM concerns. For SDXL, you’ll likely need at least 16GB VRAM for stable training—though if you’re working with lightweight adapters (like the ones in TESSRAL), you can get away with far less by freezing the backbone and only updating a tiny layer (~40MB trainable params). That’s how it runs smoothly on a 16GB-RAM M2 MacBook Air without cloud workarounds. For most use cases, 12GB+ should suffice if you’re optimizing for memory efficiency.
Most hosted versions still hit an inference endpoint. Are you trying to avoid latency or privacy leaks? I've been skeptical of "train on your own hardware" claims for years. Every project promised local training and delivered either toy results or required a 3090 minimum. TESSRAL is the first one that genuinely runs on my M2 MacBook Air with 16 GB RAM — no cloud fallback, no Colab notebooks, just
pip install tessraland it works. The architecture avoids the usual memory wall by freezing the backbone and only updating a tiny adapter layer. What surprised me most ## What are the minimum dataset requirements? - Dataset size: 30-50 images or 2-5 minutes of audio actually suffices. The adapter approach doesn't need the thousands of samples full fine-tuning demands. - No VRAM spikes: Peak memory stays under 10 GB on Apple Silicon. My 4090 desktop runs it faster but the laptop doesn't OOM. - Export format: Models save as standard safetensors + config JSON. Drop them into ComfyUI, Automatic1111, or any inference engine that accepts LoRAs — no vendor lock-in. - Sharing mechanism: You can publish a dataset hash (not the raw files) and let others train against it locally. The data never leaves your machine unless you explicitly upload it. Where it still feels rough ## Are there any significant missing features? - No built-in evaluation: You t