This live stream features an in-depth exploration and series of experiments with new AI image and video generation tools within the ComfyUI environment. The host primarily focuses on testing two new models: Stability AI's "Flux Crayer" for image generation and the "WAN 2.2" model, which is used for both video and high-resolution image creation.
Key experiments include a systematic comparison of different samplers on the Flux model and pushing its resolution capabilities far beyond documented limits, successfully generating highly detailed 9.4-megapixel images without tiling. The host then trains a new style LoRA based on the painter Mr. Reynolds for the WAN 2.2 model using the FAL.ai cloud platform. Testing this LoRA reveals a significant finding: image-trained LoRAs require a low strength (around 1.0), whereas video-trained LoRAs need a much higher strength to be effective. The stream concludes with the host being particularly impressed by WAN 2.2's ability to generate stunningly detailed still images, which he then post-processes in Photoshop for printing.
AI Image & Video Frontiers: A Deep Dive into Flux Crayer and WAN 2.2
00:08:28 Setting Up the Flux Experiment: The host adapts a complex, pre-existing ComfyUI workflow to test the new "Flux Crayer" model. He sets up a 5x5 grid to compare the outputs of 25 different sampler methods simultaneously.
00:20:03 Initial Sampler Tests: Using his "Chirona V4" character LoRA, the host begins generating the sampler grid. He observes a wide variety of results, from coherent images to artifact-heavy "cauliflower" textures from samplers like LCM when used with standard settings.
00:29:25 Life Before AI: The host takes a moment to reflect on the transformative impact of AI on creative and technical work, comparing the current moment to the invention of the mobile phone and how difficult tasks like research were "before AI."
00:45:25 Testing the "Cyber Society" Style: The host switches to his "Cyber Society" style LoRA to test with the Flux model. He is immediately impressed by the significant increase in quality, detail, and style adherence compared to previous models.
00:52:20 Pushing Resolution Limits with Flux: In a key discovery, the host experiments with generating images at resolutions far higher than the model's standard base (e.g., 1536x1536 and larger). He is surprised to find that Flux can generate these massive, 9.4-megapixel images without breaking or requiring tiling, producing extremely detailed artwork.
01:02:08 Training a LoRA for WAN 2.2: The host switches to the WAN 2.2 video model. He uses the FAL.ai cloud platform to train a new style LoRA based on the artwork of the 18th-century painter Mr. Reynolds, using a small, captioned dataset of 18 images.
01:39:10 LoRA Strength Discovery: While testing the new Reynolds LoRA for video generation, the host confirms a theory: LoRAs trained on images work best at a low strength (~1.0), while LoRAs trained on video data require a much higher strength (~6.0) to effectively apply their style. Pushing the image-trained LoRA to a high strength "cooks" the output, producing heavy artifacts.
01:52:23 Generating Images with WAN 2.2: The host uses the WAN 2.2 video model for still image generation. He compares a generation using his new Reynolds LoRA against the base model's output. He is stunned by the high quality and detail of the base model's images and confirms that his LoRA successfully applies the intended oil painting style.
02:22:17 From Digital to Print: Impressed by the results, the host takes the generated images into Adobe Photoshop. He adds film grain and other effects to enhance the canvas texture and prepares the high-resolution artwork for printing on his A3 printer.
02:39:20 Final Takeaways: The host concludes that both the Flux Crayer and WAN 2.2 models represent a huge leap in quality. He particularly emphasizes the unexpected strength of WAN 2.2 as a tool for creating incredibly detailed, high-resolution still images, making it a powerful alternative to traditional upscaling methods.