This 'Laten Space Live' session, hosted by Drift Johnson, dives into the experimental WAN 2.2 T2V low noise 14B vase in Q8 model, exploring its capabilities for both image and video generation. The video details the host's work with custom-trained Lauras, showcasing prompt enhancement via LLMs and rigorous testing of the model's integrated ControlNet (Vase). Key findings highlight WAN 2.2's impressive high-resolution image generation and consistent character learning, despite challenges with ControlNet effectiveness due to source image quality and pose model limitations, necessitating high Laura strengths. The session also touches on the rapid advancements in video AI and the potential for more accessible local model deployment.
Exploring WAN 2.2 for Image and Video Generation: Experiments and Findings
5:53 Introduction & Model Focus: Host Drift Johnson welcomes viewers to "Laten Space Live," introducing the experimental WAN 2.2 T2V low noise 14B vase in Q8 model, slated for image and video generation and training.
8:36 Custom Laura Development: The host demonstrates custom-trained Lauras, including "Star Cyber Society," showcasing how a prompt enhancer (using an LLM) significantly improves detail and definition over default prompts.
9:18 Laura Training Methodology: Details a training formula of 10-20x steps per image (e.g., 200 images for 4000 steps), with a minimum of 500 steps, and notes the flawless style learning of models like "Cyber Society 77 WAN."
15:07 Resolution & Training Fidelity: Discusses the impact of training Lauras at resolutions higher than their source images, which can sometimes lead to unexpected quality issues. Highlights the importance of "codec-safe" resolutions for optimal SDXL results.
18:25 Prompt Enhancement with LLMs: Explains the workflow for using an LLM (Ollama) to dynamically refine and blend prompts, enriching the output by incorporating specific stylistic instructions and combining elements from different source texts.
31:40 Integrated ControlNet Evaluation: Begins rigorous testing of the WAN model's built-in ControlNet (Vase) using Canny and Pose inputs. Initial Canny results are hampered by the low definition of source images.
1:33:00 Laura Strength Insights: A key finding reveals that Lauras trained on video clips require significantly higher strength settings (e.g., 5.0-6.0) to assert their style in image generation, contrasting with lower strengths (0.7-1.3) typically sufficient for image-trained Lauras.
1:47:00 Pose ControlNet Functionality: Despite challenges with input quality and PoseNet's limitations (e.g., lack of detailed face/hand detection), the ControlNet is shown to effectively influence character pose and overall composition, producing recognizable character styles like "Stella Beloo" with a laser katana.
2:05:00 WAN as a Versatile Solution: The host advocates for WAN 2.2 (especially the Vase variant) as a powerful all-in-one solution for both high-resolution image generation (producing 2K images directly without upscaling) and video generation, potentially replacing multiple specialized tools.
2:19:00 Future Accessibility & Model Optimization: Discussion shifts to advancements in AI model accessibility, including a recent 99% reduction in the size of the T5 clip encoder, which could dramatically lower VRAM requirements and make large models more viable for local deployment.
2:22:38 Concluding Remarks: The session concludes that the experimental WAN 2.2 model, when combined with carefully trained Lauras and ControlNet, shows promising capabilities, setting the stage for future streams focusing on video animation and in-painting.