New LoRA/IC‑LoRA tool updates

Two adapter tools moved forward this week: Ostris updated its AI Toolkit interface for ACE-Step 1.5 Extra Large LoRA training, and A.I.Warper published IC-LoRA outpainting tests for LTX 2.3. (github.com, docs.ltx.video) LoRA, short for Low-Rank Adaptation, is a way to fine-tune a large model by training a small add-on instead of rewriting the full model. ACE-Step 1.5 documents LoRA training for music generation with 16 gigabytes of video memory minimum and 20 gigabytes recommended for full-length songs. (github.com) Ostris’s AI Toolkit is an open-source training suite that runs as both a graphical interface and a command line tool, and its GitHub readme says it targets consumer-grade hardware. The same repository lists support for image, video, and LTX-2.3 model workflows, which puts the new ACE-Step interface work inside a broader all-in-one trainer push. (github.com) IC-LoRA, short for In-Context LoRA, does a different job. LTX documentation says it conditions video generation on reference signals such as depth maps, pose skeletons, and edge detections, giving frame-level control over structure and motion instead of only changing style. (docs.ltx.video) That distinction shows up in the official LTX comparison table: standard LoRA uses a text prompt alone, while IC-LoRA adds a control signal and is meant for structural guidance. LTX’s docs also say the two can be combined, with IC-LoRA handling layout or movement and a separate LoRA handling look or effects. (docs.ltx.video) LTX-2 itself is an open model family for synchronized audio and video generation, and the Lightricks repository says LTX-2.3 is part of that stack. The project readme describes it as a DiT-based audio-video foundation model with open access and a built-in LoRA trainer. (github.com) The outpainting side is already getting community experiments. A Hugging Face page for an LTX-2.3 22 billion-parameter IC-LoRA outpaint model describes a workflow that brightens dark regions with gamma 2.0 before generation and applies inverse gamma 0.5 afterward, using ComfyUI color-correction nodes for both steps. (huggingface.co) Official LTX model cards show how these controls are expected to plug into ComfyUI. One LTX-2.3 IC-LoRA model card tells users to load the adapter with `LTXICLoRALoaderModelOnly` and add the guide with `LTXAddVideoICLoRAGuide`, while also tracking a reference downscale factor for smaller guide inputs. (huggingface.co) What changed this week was not a new base model release, but more polish around the layer artists and tinkerers actually touch: training screens, ComfyUI graphs, and reusable workflow files. The next step, in both posts, is the same one these tools keep chasing — making custom adapters easier to train, easier to steer, and easier to reuse. (github.com, docs.ltx.video)