SYSTEM NOTICE

Auto translation by AI. Be sure, accuracy, nuances and authorial intent may not be fully reflected.
見出し画像

Trying out the FP8/GGUF version of HunyuanImage-2.1 Distilled in ComfyUI (Recommended VRAM 12-16GB)

まゆひらa

Last update 9-15-2025
* As this was just released, I may revise this article several times.
* There is also an article for the slower, standard version that is not the Distilled version..
* Related articles are summarized in the ComfyUI article list.



■ 0. Overview

▼ 0-0. Introduction

　In this article, I will explain the procedure for using "HunyuanImage-2.1 Distilled" in ComfyUI on Windows to speed up generation. We will use the FP8 and GGUF versions of the model.

　At the time of writing, there is no LoRA for HunyuanImage-2.1 Distilled, and it is provided as a Model.

　The explanation for each is in the sections below.

  • 1. Notes and resource usage

  • 2.-3. Preparation and generation of the FP8 version

  • 4.-5. Preparation and generation of the GGUF version

　Please refer to the articles below for an overview of HunyuanImage-2.1 and the procedure for standard generation without using the Distilled model.

　Please refer to the article below for how to install ComfyUI.

▼ 0-1. Related links

　These are the official models (Model, Text Encoder, VAE). They are not used in this article.

　These are the FP8 version Model, Text Encoder, and VAE used in this article.

　The FP8 version of the model is also listed on Civitai. Due to how Civitai works, the file names may differ.

These are the various GGUF models and extensions used in this article.

The GGUF version models below should also be usable (unverified).


■ 1. Notes and Resource Usage

▼ 1-1. Notes

If ComfyUI is not updated, necessary nodes may be missing. Please keep it as up-to-date as possible, as updates often include bug fixes.

To speed up subsequent generations, models loaded into VRAM may be automatically offloaded to main RAM. Therefore, it is recommended to have at least 32GB of main RAM.

▼ 1-2. VRAM Usage

In the case of the FP8 version, it can definitely be used with 16GB of VRAM. It is believed that 12GB is possible if there is enough free VRAM (dedicated GPU memory), but if it spills over into shared GPU memory, generation speed will drop significantly.

In the case of the GGUF version, it depends on the quantization type. Even so, an environment with 12–16GB of VRAM is generally suitable.

The GGUF version may be able to generate without spilling into shared GPU memory even with 8GB of VRAM by using models like "hunyuanimage2.1-v2-q3_k_m.gguf".

▼ 1-3. Generation Time

The generation times in the author's environment (GeForce RTX 5060Ti 16GB) are as follows. This does not include the time taken to load the model from disk or process the prompt. The GGUF format varies slightly depending on the model size (quantization type).

  • Distill FP8
    Approx. 54 seconds (8 steps, euler/simple, CFG=2.0)
    Approx. 40 seconds (change: weight_dtype=fp8_e4m3fn_fast)
    Approx. 21 seconds (change: weight_dtype=fp8_e4m3fn_fast, CFG=1.0)

  • Distill GGUF Q4_K_M
    Approx. 57 seconds (8 Steps, euler/simple, CFG=2.0)
    Approx. 30 seconds (Change: CFG=1.0)

For reference, the standard generation times are as follows.

  • FP8
    Approx. 115 seconds (20 Steps, euler/simple, CFG=3.5)
    Approx. 80 seconds (Change: weight_dtype=fp8_e4m3fn_fast)

  • GGUF Q4_K_M
    Approx. 135 seconds (20 Steps, euler/simple, CFG=3.5)


■ 2. Preparation (FP8 version)

▼ 2-1. Introduction

The difference between HunyuanImage Distilled and the standard version is only the model and some settings. It also uses the same Text Encoder as Qwen-Image.

▼ 2-2. Model Installation

I organize my directories by type, so I will describe it that way, but you are free to decide as you prefer.

Download each file listed at the URL below and move them to the appropriate folder. If you have already installed them, please skip this. Note that it includes the same Text Encoder used in the Qwen-Image article.

  • Comfy-Org/HunyuanImage_2.1_ComfyUI
    https://huggingface.co/Comfy-Org/HunyuanImage_2.1_ComfyUI

    • split_files/text_encoders/
      qwen_2.5_vl_7b_fp8_scaled.safetensors (8.73GB)
      byt5_small_glyphxl_fp16.safetensors (418MB)
      → models\text_encoders\hunyuan_image\

    • split_files/vae/
      hunyuan_image_2.1_vae_fp16.safetensors (773MB)
      → models\vae\hunyuan_image\


■ 3. Generation (FP8 version)

▼ 3-1. Overview

I am providing a workflow for generation, so please feel free to use it, including for modifications and redistribution.

▼ 3-2. Supplementary Notes

Since the output quality of the Distilled model is slightly inferior, parameter adjustment seems necessary. Based on the findings below, I am providing two types of workflows.

  • The person who posted the FP8 version of the model recommends a CFG value of 1.5 to 2.5.

  • The standard combination for Sampler and Scheduler is euler/simple, but I felt that changing it to dpmpp_sde/beta provided an improvement. However, dpmpp_sde is slower at generating.

  • Setting the CFG value to 1.0 significantly increases generation speed, though it appears to slightly reduce contrast.

  • Changing the Model dtype from default to fp8_e4m3fn_fast allows for faster generation at the cost of some quality.

▼ 3-3. Workflow (Normal) and Generation Example

　This is the workflow using the original settings.

　Please download the file below and drag and drop it into the ComfyUI screen.

　This is the entire workflow. I modified the existing prompt using Claude and Grok. In my environment, it takes about 54 seconds per run.

Entire workflow

　The officially listed resolutions are five types: "2048 x 2048 (1:1)", "2304 x 1792 (4:3)", "1792 x 2304 (3:4)", "2560 x 1536 (16:9)", and "1536 x 2560 (9:16)". Steps are 8, and CFG is 2.0.

Japanese anime-style illustration of a young girl with iconic anime features: large, expressive emerald green eyes, a soft, rounded face, and long, arranged golden blonde hair adorned with a delicate hair accessory - a small white ribbon bow with pearl details positioned on the side of her head. She displays a longing, wishful expression with slightly parted lips and dreamy eyes as she gazes at the mannequin. She is currently wearing a white three-quarter sleeve blouse with a charming pink ribbon bow at the collar, paired with dark blue jeans. She stands at the storefront entrance of a modern apparel store, turned toward a white fashion mannequin that displays a beautiful summer dress - a sleeveless midi-length dress with a soft lavender base color and very subtle, faded white flower patterns, featuring thin spaghetti straps and a fitted bodice with flowing A-line skirt. Her pose shows desire and admiration - one hand gently touching her cheek, her body slightly leaning forward toward the mannequin with a wistful, yearning posture. The store interior behind them features clean white walls, clothing racks along the walls, and display tables in the center showcasing folded garments, accessories, and other merchandise. The scene captures her moment of longing for the beautiful dress, rendered in a detailed and vibrant anime art style with warm, inviting lighting.

▼ 3-4. Workflow (Customized) and Generation Example

　This is the workflow where I have independently changed the settings.

  • Changed weight_dtype in Load Diffusion Model (default → fp8_e4m3fn_fast)

  • Changed sampler/scheduler (euler/simple → dpmpp_sde/beta)

  • Changed cfg to 1.0 (Negative Prompt is disabled)

　Please download the file below and drag and drop it into the ComfyUI screen.

　This is the entire workflow. In my environment, it takes about 41 seconds per run.

Entire workflow

　I feel that changing to dpmpp_sde/beta improves the quality. However, since dpmpp_sde is slow, I have increased the speed by changing the weight_dtype and CFG.

Prompt is the same as above


■ 4. Preparation (GGUF version)

▼ 4-1. Introduction

　Since the explanation follows the same flow as the FP8 version, some content may overlap.

　The only differences between HunyuanImage Distilled and the standard version are the model and some settings. It also uses the same Text Encoder as Qwen-Image.

▼ 4-2. Setting up the models

　I organize my directories by type, so I will describe it that way, but you are free to decide as you prefer.

　Download each file listed at the URL below and move them to the appropriate folders. If you have already set them up, please skip this. The Q4_K_M model provides sufficient quality, but you can adjust it up or down while monitoring VRAM usage and quality. Note that it includes the same Text Encoder used in the Qwen-Image article.

▼ 4-3. About required extensions

When loading the GGUF version workflow posted in this article, if extensions are missing, a "Some Nodes Are Missing" message will appear. In that case, you can install them using the ComfyUI Manager features.

For instructions on how to install extensions, please refer to the "Automatic Installation Procedure" or "Manual Installation Procedure" in the article below.

The workflow in this article uses the following extensions.


■ 5. Generation (GGUF version)

▼ 5-1. Overview

Since the explanation follows the same flow as the FP8 version, some content may overlap.

I am providing a workflow for generation, so please feel free to use it, including for modifications and redistribution.

▼ 5-2. Supplementary notes

Since the output quality of Distilled models is slightly inferior, parameter adjustment seems necessary. Based on the findings below, I am providing two types of workflows.

  • The person who posted the FP8 version of the model recommends a CFG value of 1.5 to 2.5.

  • The combination of Sampler and Scheduler is euler/simple by default, and I felt it improved when changed to dpmpp_sde/beta. However, dpmpp_sde has a slower generation speed.

  • Setting the CFG value to 1.0 significantly increases generation speed, but it appears that the contrast decreases slightly.

  • Changing the model dtype from default to fp8_e4m3fn_fast can increase generation speed at the cost of some quality.

There is a point to note when using this. If you change the prompt from the second time onwards, execute "Unload Models" before starting generation. If data remains in the dedicated GPU memory (VRAM), the Text Encoder may load and spill over into shared GPU memory, which can significantly increase prompt processing time.

Execute "Unload Models" as appropriate

▼ 5-3. Workflow (Normal) and generation examples

This is the workflow using the original settings.

Please download the file below and drag and drop it onto the ComfyUI screen.

This is the entire workflow. I modified the existing prompt using Claude and Grok. In my environment, it takes about 57 seconds per run.

Entire workflow

The officially listed resolutions are five types: "2048 x 2048 (1:1)", "2304 x 1792 (4:3)", "1792 x 2304 (3:4)", "2560 x 1536 (16:9)", and "1536 x 2560 (9:16)". Steps are 8, and CFG is 2.0.

Japanese anime-style illustration of a young girl with iconic anime features: large, expressive emerald green eyes, a soft, rounded face, and long, arranged golden blonde hair adorned with a delicate hair accessory - a small white ribbon bow with pearl details positioned on the side of her head. She displays a longing, wishful expression with slightly parted lips and dreamy eyes as she gazes at the mannequin. She is currently wearing a white three-quarter sleeve blouse with a charming pink ribbon bow at the collar, paired with dark blue jeans. She stands at the storefront entrance of a modern apparel store, turned toward a white fashion mannequin that displays a beautiful summer dress - a sleeveless midi-length dress with a soft lavender base color and very subtle, faded white flower patterns, featuring thin spaghetti straps and a fitted bodice with flowing A-line skirt. Her pose shows desire and admiration - one hand gently touching her cheek, her body slightly leaning forward toward the mannequin with a wistful, yearning posture. The store interior behind them features clean white walls, clothing racks along the walls, and display tables in the center showcasing folded garments, accessories, and other merchandise. The scene captures her moment of longing for the beautiful dress, rendered in a detailed and vibrant anime art style with warm, inviting lighting.

▼ 5-4. Workflow (Custom) and Generation Examples

This is the workflow where I have independently changed the settings.

  • Changed sampler/scheduler (euler/simple -> dpmpp_sde/beta)

  • Changed CFG to 1.0 (Negative Prompt is disabled)

Please download the file below and drag and drop it onto the ComfyUI screen.

This is the entire workflow. In my environment, it takes about 55 seconds per run.

Entire workflow

I feel that the quality is improved by changing to dpmpp_sde/beta. However, since dpmpp_sde is slow, I have increased the speed by changing weight_dtype and CFG.

Prompt is the same as above


■ 6. Bonus

▼ 6-1. Bonus Images

Here are some images generated with HunyuanImage-2.1. The workflow uses the custom version from 3-4. The prompts are the ones posted in the regular version article, passed through Hunyuan-PromptEnhancer again.

A young girl with characteristic anime features is seated at a rustic wooden table on a charming countryside cafe terrace. She possesses a softly rounded face, large and expressive brown eyes, and long, dark brown hair arranged neatly. A shy, bashful smile graces her lips, complemented by a subtle blush on her cheeks. She strikes a modest pose, with one hand resting gently against her cheek. Her attire consists of a white short-sleeved blouse with a delicate lace collar, a navy floral midi skirt, white socks, and brown loafers. On the table before her sits a coffee cup. A large speech bubble is positioned above her head, containing the Japanese text "一緒に休も？" with a line break. The background depicts a serene, remote mountain village, complete with quaint stone cottages, winding dirt paths, and vibrant wildflower meadows. Distant rolling hills are visible under a sky illuminated by the warm golden light of the hour, which casts dappled lighting across the entire scene. The image is presented as a Japanese anime-style illustration, characterized by a soft earth-tone color palette and a detailed, vibrant art style.
A joyful yet shy young Japanese woman is captured in a close-up portrait, performing elegantly while the night sky illuminates the scene. The central figure gazes directly forward, her expression a mix of delight and bashfulness. Her long, flowing silver-gray hair frames her face, meticulously styled with small double chignones, each tied with a vibrant red ribbon. She wears a flowing white dress, its fabric gently rippling from a subtle wind. She is gracefully playing a violin, with the instrument positioned near her shoulder. Ethereally floating around her are shimmering musical notes, conveying a sense of serene performance. The setting is at night, dominated by a large, luminous full moon in a deep blue sky dotted with scattered stars. In the far distance, a range of mountains is silhouetted against the sky. The soft moonlight casts a gentle glow on her figure and the immediate surroundings, which include Japanese silver grass (susuki) with their distinctive bushy seed heads, all of which are affected by the gentle wind creating soft movement. The overall image is rendered in a detailed and emotive anime art style.
An illustration captures a crisp autumn day in a park filled with maple trees, with a cheerful young girl serving as the central figure. In the foreground, the young girl stands with fair skin, large, expressive eyes, and a joyful expression. Her auburn hair is styled in intricate braids. She wears a cream-colored knit sweater dress and a bright red scarf around her neck. She holds a comically large wooden sign, on which the text "Welcome to HunyuanImage-2.1!" is written in an adorable, wobbly script. A cascade of colorful maple leaves falls from its sides, appearing as if caught in a gentle breeze. Behind her, a winding stone path is covered in a thick layer of fallen red and gold leaves. In the middle distance, other visitors are depicted taking photographs, their figures slightly blurred to create a sense of depth. The background is dominated by a dense grove of maple trees, their brilliant red and gold foliage forming a rich canopy. The entire scene is illuminated by warm, golden afternoon sunlight filtering through the leaves, casting long shadows and creating a magical, glowing atmosphere with visible light particles dancing in the air. The artwork is presented as a high-quality digital illustration in a classic anime style.
The scene is set within a grand, ornate concert hall, focusing on a young girl seated attentively at a grand piano. The central figure is a young girl with light gray hair neatly styled in two pigtails that frame her face. She wears a white ribbon blouse and a sky blue skirt that is decorated with a pattern of small white floral designs. Her delicate fingers are positioned on the ivory keys of the piano, poised as if she is about to play. Perched atop the polished, glossy black surface of the grand piano is an elegant orange tabby cat, its coat featuring distinct, darker orange stripes. In the background, the vast hall is defined by a high, vaulted ceiling, which is illuminated by the warm, golden glow cast by large, ornate crystal chandeliers. Far below, the audience occupies the rows of deep red velvet seats, whose forms are rendered as dim, indistinct silhouettes in the hall's recesses of light. The overall artistic style is a soft watercolor painting, influenced by anime character design, notable for its large, expressive eyes and delicate line work, which captures an intimate performance moment.
A young girl with smooth, fair skin is seated at a dessert buffet table within a bright and luxurious hotel restaurant. Her long, dark brown hair is styled in a high ponytail, accessorized with a simple red bow. Her facial features are characterized by large, warm brown eyes, a small nose, and a gentle, natural smile. She is wearing a pastel pink summer dress featuring delicate white lace trim along the neckline. In front of her, a white ceramic plate holds a modest assortment of desserts, including a slice of strawberry shortcake with visible layers, a small scoop of rich chocolate mousse, a couple of colorful fruit tarts, and a few pastel-colored macarons. Adjacent to her plate, a tall, clear glass contains iced tea, complete with visible ice cubes and a slice of lemon submerged within the liquid. The background is softly blurred, showing the indistinct forms of other dessert stations and the upscale interior of the restaurant, creating a shallow depth of field. The entire scene is bathed in soft, warm lighting that casts gentle highlights and creates a cozy, inviting atmosphere. This image presents a photorealistic photography style.
A slightly elevated, close-up shot captures a young Japanese woman in mid-stride while hiking on a sun-drenched mountain trail, presented from a dynamic, low-angle perspective. The central figure is the woman, her upper body and walking motion occupying the majority of the frame. She has joyful and energetic facial expressions, with her eyes looking forward, and her hair is pulled back into a high ponytail, partially covered by a simple cap or headband. She is dressed in light summer hiking gear, consisting of a colorful, short-sleeved tank top or a light, breathable hiking shirt, paired with shorts. In one hand, she grips a trekking pole, using it for balance as she moves along the path. A smaller daypack is visible, strapped to her side. The trail itself is narrow and unpaved, flanked by the lush, vibrant green vegetation of a mountainous region during summer. In the background, layers of forested slopes and distant mountain peaks are visible under a bright, clear sky. The entire scene is illuminated by bright, natural sunlight. This image presents a photography style, characterized by a shallow depth of field that keeps the subject sharp against a softly blurred background.
A professional businessman is captured in a slightly elevated, close-up shot, frozen in mid-stride as he walks with purpose across a bustling New York City street. The central figure is a man with a determined expression, his face showing confidence as he moves forward. He is dressed in a sharp, well-tailored business suit of a deep navy or charcoal gray color, worn over a crisp dress shirt and a complementary tie. In one hand, he carries a dark leather briefcase or a sleek laptop bag. The camera angle is a low diagonal, focusing on his upper body and capturing his dynamic walking motion, which creates a sense of forward momentum. He moves along a wide city sidewalk, where the pavement is lightly textured and reflective. In the background, the iconic urban scenery of New York City is rendered with a shallow depth of field, creating a soft, blurred bokeh effect. Recognizable shapes include towering skyscrapers, the vibrant yellow of a taxi cabs, and the general bustle of city life. Natural daytime lighting illuminates the scene, casting soft shadows and highlighting the textures of his suit and the surrounding environment. This image presents a photography style.
A lively scene depicts a group of diverse plush dolls engaged in conversation within a brightly lit school hallway. In the foreground, three central dolls form an animated group. The doll on the left is a girl with long brown hair styled in pigtails, wearing a classic navy blue and white sailor-style uniform with a red ribbon. She holds a miniature textbook in one hand and a tiny pencil case in the other, her fabric mouth open as if speaking. The central doll has short black hair and wears a light pink sailor uniform with a matching bow, holding a cute pink eraser and tilting her head curiously. To the right, a third doll with blonde hair in a neat bun is dressed in a blue-and-white uniform with a yellow ribbon, gesturing with one hand as she listens. The background features the hallway setting, including a row of colorful lockers lining one wall and large windows on the other, which allow warm, cheerful light to stream in. A bulletin board, scaled down to fit the doll's height, is mounted on the wall between the lockers. The entire scene is rendered in a soft-focus, 3D digital art style, emphasizing the plush texture and fuzzy details of the toys.
A surreal and cinematic landscape unfolds on an alien world, where realism and fantasy converge. In the foreground, unknown beings with elongated limbs and skin that has a glossy, liquid-like quality move gracefully through the dreamscape. They glide over shimmering metallic vegetation that shifts between organic, leaf-like plants and sharp, geometric crystal formations. The middle ground is dominated by a series of massive, spiraling towers that ascend towards the sky, their metallic surfaces twisting into complex, fantastical shapes. Above the towers, strange bioluminescent creatures, resembling living jellyfish with translucent, bell-shaped bodies and long, delicate tentacles, drift slowly through the air, emitting a soft, ethereal glow. The sky itself is a deep, cosmic expanse filled with multiple moons of varying sizes and colors, which are surrounded by vibrant aurora streams that swirl in shades of green, purple, and blue. The overall lighting shifts between warm, golden ethereal glows and cool, blue-toned illuminations, casting dramatic shadows and highlights across the entire scene. This image is rendered in a highly detailed, surreal cinematic style, reminiscent of a still from an epic science-fiction film.


■ 7. Other

　Please follow the menu to find other articles I have written.

　There is also a list of articles limited to ComfyUI. as well.

　If you have any questions regarding the articles, please contact @riddi0908 on X.

いいなと思ったら応援しよう！