This video delves into advanced AI image generation workflows, specifically focusing on building and comparing custom prompt enhancement systems within ComfyUI. The speaker demonstrates how different Large Language Models (LLMs) – including GPT-5, Gemini, and Sonnet (Claude Opus 4.1) – can be leveraged to craft "system prompts" that guide a base LLM (Quen 2.5VL) in generating enhanced image prompts. Key experiments involve integrating custom LORAs (Chirona, Cyber Society, and a newly trained "Thora" character model) into these enhancement systems for both text-to-image and image-to-image generation.
The demonstration reveals significant variations in LLM performance, with Sonnet and Gemini consistently creating more effective and style-aligned enhancement prompts compared to GPT-5, which initially struggled with fundamental aspects like LORA trigger words and desired art styles. Despite some "creative" failures (e.g., a "cat centaur"), the refined prompt enhancers, particularly those authored by Sonnet and Gemini, achieve remarkably consistent character rendition, intricate details (like firearms and eyes), and effective style blending, often surpassing the quality of the original LORA training data. The speaker highlights the efficiency of image-based testing for faster feedback before tackling more resource-intensive video generation, which is planned for future sessions.
Summarizing the AI Image Generation Journey: Prompt Enhancers and LLM Comparisons
3:36 Project Introduction: The video introduces a project to create and share advanced image generation workflows on GitHub, focusing on "ENH modules" (prompt enhancement systems) within ComfyUI.
4:38 Image-to-Image Comparisons: The speaker begins by comparing traditional image-to-image methods (VAE encoder) with newer "WAN Image to Video" nodes that incorporate Clip Vision, noting that the latter can sometimes behave more like text-to-image, losing fidelity to the original input image's composition.
5:01 Workflow Breakdown: The core workflow uses WAN 2.2, WAN 2.1 VAE, and Olama running Quen 2.5VL for prompt input. Custom LORAs like "Cyber Society" are integrated for style.
8:22 LORA Customization: The goal is to create customized image-to-image prompt enhancement systems that are aligned with specific LORAs, such as "Cyber Society" or a combined "Chirona" and "Cyber Society" style.
13:02 Crafting System Prompts: The speaker manually refines system prompts to ensure they guide the enhancement process to incorporate LORA-specific elements like "consistent character models" and "trigger-based style models," even experimenting with "segmentational separation."
17:15 Initial Image Generations: Early results showcase the ability of enhancers to blend styles (e.g., Chirona with Cyber Society features), but also highlight challenges in maintaining exact character pose or producing unexpected, "horrific" AI creations like a "cat centaur" (1:14:03).
45:30 Generating Test Images: The speaker generates various images using different prompts (e.g., "witch's cat") to test the enhancement systems, showcasing how they can transform simple inputs into highly detailed outputs, often improving aspects like backgrounds and overall artistic quality.
1:01:00 Introducing "Thora" LORA: A newly trained "Thora" character LORA (v0) is introduced, trained with both character images and "regularization images" of sci-fi scenes to improve background generation.
1:03:00 "Coffee Time" Example: A quick demonstration of WAN's ability to consistently render text within images ("Coffee Time" signs) across different styles and transformations, illustrating its robust generation capabilities.
1:30:00 ComfyUI Crashes: An attempt to run the image-to-video motion enhancement system results in ComfyUI crashing due to memory issues, prompting a restart and a focus on image-based testing for efficiency.
1:37:00 Workflow Release Plan: The speaker outlines plans to release the various prompt enhancement system prompts as text files on GitHub, emphasizing a philosophy of free sharing for community benefit.
1:53:00 LLM System Prompt Comparison: A core experiment involves having different LLMs (Gemini, ChatGPT-5, Sonnet/Claude Opus 4.1) generate system prompts for the "Thora" LORA. The results are then used to enhance prompts for image generation with Quen 2.5VL.
2:15:00 "Thora" LORA Performance: The "Thora" LORA performs exceptionally well, producing highly detailed images with consistent character features, impressive firearms, and good anatomy, often surpassing the quality of the original training data.
Sonnet 4 & Gemini: Consistently generate effective system prompts that align well with the LORA's style and character, producing high-quality, stylized images.
ChatGPT-5 (Initial Attempt): Performs "embarrassingly bad," failing to include the LORA's trigger word and producing generic, un-stylized images.
ChatGPT-5 (Second Attempt): Improves after explicit re-instruction ("beating it over the head"), but still struggles to fully capture the "anime" style compared to Sonnet and Gemini.
2:57:00 "Uncensored" Models: The speaker highlights that "uncensored" models are beneficial not for explicit content, but because they have a broader understanding of anatomy, leading to more realistic and detailed character rendering.
3:06:00 Ongoing Enhancer Testing: Further tests with "sneaking" poses and "closeup" shots demonstrate the varying artistic interpretations and quality across different LLM-generated enhancers.
3:18:00 Character Consistency Observation: A drawback of WAN for character consistency is noted: if a LORA isn't trained with multiple characters, all characters generated tend to resemble the single LORA character.
3:40:00 Final LLM Ranking: For this specific task, Sonnet 4 is deemed the winner for system prompt generation, followed by Gemini in second place, and GPT-5 in third.
3:45:00 Conclusion & Next Steps: The session concludes with a strong endorsement of the quality achieved in image generation and an anticipation of exploring video generation with these enhanced workflows in the next stream.