HiDream i1 vs Flux Dev | ¿Cuál genera mejores imágenes con IA? + Comparativa y Upscaling en ComfyUI

Introduction to Highdream i1 vs. Flux Dep

Overview of the Battle

• Mauricio Perdomo introduces the anticipated comparison between the new Highdream i1 image generation model and the established Flux Dep model.

• The analysis will cover quality, efficiency, speed, and available quantizations of both models.

Highdream i1 Features

• Highdream i1 is a free AI image generation model compatible with Confi UI and comes in three versions: Fast, Dep, and Full.

• Each version caters to different performance needs:

• Fast: Lightweight for quick generation with lower resource consumption.

• Dep: Balanced option between quality and speed.

• Full: Heavyweight for maximum detail, recommended for powerful GPUs.

Quantization Options Explained

Model Sizes and Requirements

• The FP16 version weighs 34.2 GB; requires a GPU with at least 12 or 16 GB of VRAM for optimal performance.

• The FP8 version is compressed to 17.1 GB, allowing it to run on GPUs with 8 or 12 GB of VRAM but may have longer image generation times.

Comparison Setup

• A flow in Confi UI will compare all three versions (Fast, Dep, Full), including their respective FP8 compressions against Flux Dep's Goof Q8 format.

Downloading Models from Hugin Face

Steps to Download

• Users are directed to download models from Hugin Face under the Confork profile for Highdream i1.

• Four text encoders are required by default for Highdream:

• Clip G Hydream

• Clip L Hydream

• Lama model (3.1 billion)

• T5 XXXL FP8

Testing Different Formats

Available Formats

• Users can find specific repositories on City 96’s Hugin Face profile for various versions of Highdream.

• Recommendations based on GPU capabilities:

• Less than 8 GB RAM: Test Q4KS or Q4KM formats.

• At least 8 GB RAM: Test Q8 formats which may match FP16 quality while being lighter (18.7 GB).

Preparing for Testing in Confi UI

File Organization Tips

• Models should be stored in designated folders within Confi UI:

• FP8/FP16 models in Conffi UI Models/Diffusion Models.

• Goof formats (Q8/Q4) in Models Unit.

Application Updates Required

• Before testing, users must update their application and custom nodes through options provided within Confi UI to ensure compatibility and performance.

Image Generation Workflow and Model Comparisons

Setting Up the Image Generation Process

• The process requires a restart after updating configurations. A general group is created for generating images across different models using a fixed seed for one-to-one comparisons.

• An empty latent image is configured, which will be input into the Campler. The BAUE used is consistent with previous flux usage, along with a quadruple clip node necessary for High Dream.

• Positive and negative prompts are set up; only full versions of Highdam require a negative prompt, while fast and dep versions also need it similar to flux.

Model Configuration Details

• According to Highdere's official documentation, the full model requires a sampling chief of TR with 50 steps, using Unipaler Simple sampler and CFG of 5.0 to account for the negative prompt.

• For Highdream Dep, it recommends a sampling chief of 6.0 or 28 steps for faster processing, utilizing LCM sampler with normal scheduler and CFG of 1 since it does not consider the negative prompt.

• Hydrinam Fast is noted as the quickest option requiring only 16 steps at modeling chief of 3.0, maintaining LCM sampler and normal scheduler without accounting for the negative prompt (CFGD at 1.0).

Generating Images Across Different Models

• Each section in the workflow varies based on recommended configurations per model variation; initial settings include Hydream Fast FP8 parameters like modeling of 3, recommended steps, CFG of 1, etc.

• Additional nodes are included to play sound and clear cache memory; similar workflows are repeated for Hydream Fast Q8 using Unit Loader GGUF while keeping other configurations constant.

• For High Dream Dep model configuration: modeling set at 6 requiring more steps (28), maintaining CFG at 1 due to no consideration for negative prompts alongside testing DEP Q8 and Q4 models.

Full Model Recommendations

• The flow setup includes recommendations for Hydre Full model: modeling TR with suggested 50 steps that may significantly increase image generation time depending on GPU capabilities.

• It’s advised that users should have over 16 GB video RAM when working with full models due to potential long generation times (up to half an hour).

• For full models' settings: CFG set at 5.0 considering negative prompts; recommended samplers include UniPC with simple scheduler while repeating processes for Hydrin Full Q8 and Q4 variations.

Comparing Image Outputs from Different Models

• The final workflow incorporates image generation via flux depo allowing comparison between two models’ outputs; recent evaluations show High Dream outperforming Flux 1.1 Pro in quality assessments.

• A specific scenario is generated where an image depicts a woman walking and greeting attendees; both positive and negative prompts are utilized during this process focusing on Hydre fast FP8 execution first.

• Initial results from Hydream Fast FP8 show promising outcomes despite common issues in hand generation across various models; this version successfully maintains accurate finger representation.

Evaluating Generated Images

• Subsequent variations generated include Hydream fast Q8 and Q4 leading to comparative analysis among three images produced under different compression formats (FP8, Q4, Q8).

High Dream Dep vs. Flux Dep: A Comparative Analysis

Introduction to High Dream Dep

• The discussion begins with an introduction to the "High Dream Dep" model, highlighting its potential for generating surprising results.

• The speaker mentions disabling previous versions of "Hydream Fast" to focus on the "Hydream Dep FP8" model for image generation.

Image Generation Process

• Confirmation of the correct model selection is made, utilizing a sampling method set at six and increasing steps to 28.

• After generating three images in FP8 format, it is noted that one image had issues with hand generation, suggesting that different seeds might yield better results.

Model Comparisons

• A comparison between "Highdream Fast Q8" and "Highdream Dep" reveals that while both models have strengths, the latter shows superior color quality despite some errors.

• Transitioning to "Flux Dep," the same prompt is used without additional modifications; this model's speed in generating images is highlighted as a significant advantage.

Performance Insights

• The speaker notes that "Flux Q8" generated images faster than expected but acknowledges that "High Dream" interprets prompts more effectively due to its advanced architecture.

• Three images are displayed side by side: Hydream Fast on the left, Flux Dep in the center, and Hydre Dep on the right. The need for precise prompts for optimal results from Flux is emphasized.

Further Comparisons and Results

• Previous attempts using similar prompts yielded varying results; notably, FP8 compression produced better outcomes in hand representation.

• A detailed comparison between variations of models indicates that while higher quality models like Hydream Full can enhance detail, they struggle with specific elements like hand generation.

Conclusion on Style Differences

• Observations about stylistic differences show that while Flux may produce bolder interpretations (e.g., dresses with transparencies), Highdream maintains a more traditional approach aligned with prompt expectations.

• An exploration into anime-style images demonstrates how each model performs differently regarding background design and overall composition quality.

Animation Model Comparison

Overview of Hydream Models

• The discussion begins with a comparison of three models from Hydream Fast, highlighting preferences for the FP8 or Q8 models based on animation style.

• A focus on the Highdream Dep model reveals improvements in lighting compared to previous versions, emphasizing detail and image quality in larger models like Q8.

Realism and Execution Time

• The full model's advantage lies in its realism; however, it requires significantly more computational steps (50 vs. 28 for High Dream Dep), impacting execution time.

• Users with high-end GPUs (RTX 3090 or 4090) can efficiently run the full model, while those with less powerful machines may prefer High Dream Dep due to its better quality-to-step ratio.

Image Comparisons

• A side-by-side comparison of Flux Dep and Highdream Dep shows both producing appealing images, though Flux offers warmer colors.

• In generating an image of a paella, only the Hydre Full model accurately depicted four handles on the pan; other models struggled with this detail but performed well overall.

Performance Insights

• The results indicate that while Flux has strengths in realism, Highdream is positioned to potentially surpass current Flux models if additional tools are developed.

• Feedback is solicited from viewers regarding their preferred model based on visual comparisons presented.

Image Generation Techniques

New Workflow for Image Generation

• Introduction of a new workflow allows automatic extraction of prompts from uploaded images for enhanced image generation capabilities.

• Adjustments made to latent images enable efficient image-to-image transformations using the Highdream Dep model with reduced step counts (as low as five).

Style Transformation Capabilities

• Emphasis on configuring Dinois values to control creativity levels during image transformations; users can experiment between values like 0.3 to 0.5 for varying effects.

• An example transformation aims to convert an image into a Pixar-style animation by combining prompts effectively within the workflow.

Results and Future Applications

• The generated output demonstrates successful style adaptation while maintaining core elements of the original scene; adjustments were made without losing essential details.

High Dream Depth: Image Generation Comparison

Overview of High Dream Depth Model

• The video discusses the results generated by the High Dream Depth model in image-to-image workflows, showcasing its capabilities and performance.

• A comparison is made between three versions of the Highdream model: Fast, FP8, Q4, and Q8, highlighting their differences in output quality and processing speed.

• Additional comparisons are drawn with the Flux Dep model to evaluate how these models perform against each other in generating images.

Practical Application and Resources

• A simple workflow for generating images using the Highdream Dep model is demonstrated, providing viewers with practical insights into its application.

• Viewers are encouraged to download the workflows for Confi to conduct their own tests and comparisons of these new models; a link is provided in the description for easy access.