Verboa Image 1.0

Content warning. This model generates photorealistic adult sexual content on request. The images on this page are explicit. The training data is adult studio photography, screened for age as described under Safety. The model is released under a license whose use restrictions are part of the terms (see License), and it had a weight-level safety pass before release. Please do not use it if you are under 18 or where adult content is illegal.

Video: turn any Verboa still into a clip with sound with our free MiniMax H3 workflows for ComfyUI, in Verboa-Prompt-Writer/Verboa-H3-Video (also on CivitAI).

Verboa Image 1.0 is an 8B photorealistic text-to-image model for adults. It is a full fine-tune of ERNIE-Image, Baidu's single-stream diffusion transformer, on about 300,000 captioned images, roughly 250,000 professional adult photographs and 50,000 openly available general images made fast by transplanting ERNIE-Image-Turbo's few-step ability onto it. It finishes an image in 8 steps. Our goal for this project was simple: photographic realism for adult subjects, from prompts written the way people actually write them.

Highlights

• Eight steps. 8 steps at guidance 2, 16 forward passes per image. Fast enough for a laptop GPU, usable on a Mac.
• Plain-language prompts. Every training caption is the prompt a particular person would type: one to a hundred words, slang to clinical, casual to formal.
• A prompt writer, if you want one. Verboa Prompt Writer, a small companion model trained on the same captions, writes the prompt for you in ComfyUI: set a few fields, or let it surprise you.
• Realism first. Skin, light, lenses, real rooms. The adult set is professional studio photography that passed an AI-image detector and watermark removal for quality.
• Explicit vocabulary. Solo, partnered, same-sex, groups; the acts and positions people ask for, captioning the way they ask for them.
• Still a general model. The 50,000 general images (photos, art and web pages) are there to keep ERNIE-Image's instruction following, text rendering and scene composition.
• Open weights, commercial use allowed under CreativeML Open RAIL++-M with one added use restriction (see License); the use restrictions travel with every copy and derivative.

Showcase

The three panels at the top of this card, prompt by prompt: portraits and nudes (1 to 6), explicit (7 to 12), and men and couples (13 to 18). Every image: Verboa Image 1.0, 8 steps, guidance 2.0, 848×1264, empty negative prompt, no prompt enhancer, rendered in diffusers with fp8 weights. Each was chosen from four seeds of its prompt (92301 to 92304) and is shown as rendered, with no upscaling, inpainting or retouching. The panels scale the tiles down and crop some to fit their shape.

Model details

How it was made

1. Fine-tune. Every transformer weight of ERNIE-Image was trained on the dataset below for 11,338 updates. That checkpoint samples well at 50 steps, but it won't be released since we never gave it the safety edit in step 3.
2. Speed. ERNIE-Image ships a distilled Turbo variant. The difference between Turbo and ERNIE-Image is a large, high-rank, full-weight change that buys few-step sampling. Adding 0.85 of it to the fine-tune transplants that ability without a distillation run of our own.
3. Safety edit, on the transplanted weights. An erasure applied before the transplant was measurably undone by it: the Turbo delta is a big change computed between two models that both still have the concept, so it swamps a small edit. The safety edit was therefore fitted directly to the transplanted weights and merged last.

Quick start

diffusers

pip install -U diffusers transformers accelerate safetensors

import torch
from diffusers import ErnieImagePipeline

pipe = ErnieImagePipeline.from_pretrained(
    "verboa/Verboa-Image-1.0",
    torch_dtype=torch.bfloat16,
    pe=None, pe_tokenizer=None,          # no prompt enhancer
).to("cuda")

image = pipe(
    prompt="a cute 27 yr old woman with long black hair sitting naked on a bed with blue sheets",
    negative_prompt="",
    height=1264, width=848,
    num_inference_steps=8,
    guidance_scale=2.0,
    use_pe=False,                        # required
    generator=torch.Generator("cuda").manual_seed(0),
).images[0]
image.save("verboa.png")

pipe.enable_model_cpu_offload() fits the bf16 model on a 24 GB card. On 24 GB the fp8 file is faster, because it fits resident without offloading, use it through ComfyUI.

ComfyUI

ERNIE-Image is supported natively, so the image model needs no custom nodes. Put the files in place and load workflows/verboa-image-1.0.json (NVIDIA) or workflows/verboa-image-1.0-mac.json (Apple Silicon):

The workflow opens in App Mode with prompt, size, seed, steps and cfg as the only controls. If you build your own graph, these two settings matter: a ModelSamplingSD3 node at shift 4.0 with the simple scheduler (together they reproduce the diffusers pipeline's schedule exactly; ComfyUI's default shift for ERNIE is 3.0), and an empty CLIPTextEncode as the negative, not ConditioningZeroOut, the model was trained with the empty prompt as its negative. With both, ComfyUI and diffusers render an identical picture.

Or let the Prompt Writer write the prompt. Verboa Prompt Writer comes with one small custom node and its own versions of these workflows. The node has three modes: Let Me Choose (set people, age, place, how explicit, voice, length and a few more, or leave them open), Surprise Me (it picks for you) and Write My Own (your own prompt, as above). Each run writes a new prompt and renders it. Files and setup are on its card.

mflux (Apple Silicon)

mflux-generate-ernie-image --model verboa/Verboa-Image-1.0-mflux-8bit --steps 8 --guidance 2 --height 1264 --width 848 \
  --prompt "a cute 27 yr old woman with long black hair sitting naked on a bed with blue sheets"

Set steps and guidance yourself: mflux's ERNIE-Image default is guidance 4.0, and its Turbo command accepts only 1.0.

Hardware

Downloads

Every file is the same model at a different precision. The repos are gated: open the model page while logged in to Hugging Face and accept the terms, then run hf auth login on the machine that you want to run it on (diffusers and mflux use that login).

• NVIDIA RTX 40 and 50: fp8. On these cards ComfyUI computes in fp8 as well as storing the weights in it.
• RTX 50 with little memory: NVFP4, which ComfyUI runs with 4-bit activations on these cards.
• Older NVIDIA cards: fp8 and NVFP4 still load but are dequantized as they run; int8, which ComfyUI always runs weight-only, and the GGUF files are the alternatives.
• Mac: bf16 or int8 in ComfyUI (Apple Silicon has no fp8); int8 will draw the same picture in half the memory. GGUF runs too, but more slowly. Or mflux.

Prompting

Write what you want the way you would say it. The captions the model learned from were generated to look like real user prompts, drawn from a grid of personas:

• Length: from two words ('nude woman') to a 140-word paragraph. Short prompts give the model freedom; long ones are followed closely. A one-sentence prompt of 15 to 40 words is the sweet spot.
• Voice: slang, casual, neutral, formal and clinical vocabularies were all trained (28 / 32 / 15 / 12 / 13% of captions). "Pussy", "vulva" and "genitals" all work and mean the same thing, same with "dick", "cock" and "penis".
• Ages are written the way people write them: "in her twenties", "a woman in her thirties", "18 year old", "mature". Every caption age is an adult age (18 or older); about a fifth of the adult captions use coded terms such as "milf", "college girl", or "barely legal" instead of a number.
• Negations work inline ("no tattoos", "no jewelry"), because 12% of captions carried one that was true, not required.
• The core of an explicit prompt is the person (hair, build, age), what they are wearing or not, and what they are doing. Setting, light, camera and mood are extras. A setting alone does not imply sex: "a bedroom" gives a bedroom.
• Sampler: 8 steps, guidance 2.0, empty negative prompt. Do not drop guidance to 1.0: at 8 steps and CFG 1, earlier builds of this model often failed to draw a face at all, and every measurement in this card was taken at CFG 2. If faces matter for a prompt, more steps help.
• Do not run the ERNIE prompt enhancer, and do not use tag-style prompts from other model families ("1girl, masterpiece, best quality"); neither was trained.

Examples in four voices, all valid:

Tag: nude woman, 20s, painting at easel, loft art studio, exposed brick, paint-stained forearms, afternoon sunlight, dust in light beams, candid, 35mm film grain
Casual: hot nude woman in her twenties painting in her apartment, paint all over her arms, afternoon sun coming through the big windows
Photography: Candid 35mm photograph of a nude woman in her 20s standing at an easel in a converted loft studio, brush raised, streaks of ochre paint on her forearms and hip. Low afternoon sun through tall factory windows, dust in the beams, exposed brick, canvases stacked against the wall. Portra 400 tones, shallow depth of field, natural skin texture.
Storyteller: Late afternoon in an old brick loft. A woman in her twenties stands naked at her easel, lost in the canvas, one brush between her teeth and another in her hand, blue and ochre drying on her wrists. The sun comes in low through the tall windows and turns the dust gold, and she hasn't noticed the light crawl across the floor for an hour.

Rather not write one? Verboa Prompt Writer learned from these same captions and writes prompts like them for you in ComfyUI.

Training

The adult set is 250,000 photographs from professional studio and studio-aggregator sources, 16,253 of them from two dedicated male sources; the general set is 50,000 images from six public datasets. Every image got one caption written as a user prompt (below), and the fine-tune ran 11,338 updates at batch 64 on an NVIDIA B200, about 82 hours.

Captions

One caption per image, written by Qwen3.6-35B-A3B (NVFP4, vLLM) from the image at 1024 px plus a per-image persona drawn from the image's hash: who is typing, what they care about, how long, in what format, which vocabulary. The result is a prompt distribution, not a description distribution. Every explicit caption carries the core (person, undress, act); 12% of focused captions add a negation to the prompt itself. A cleanup pass stripped non-description text and re-captioned model refusals; an age pass rewrote age words into adult forms and dropped ages where no face made one readable.

Evaluation

There are no public benchmark numbers for this model. You may run the model through your own benchmarks if you like, would love to see them! GenEval-style scores measure things the fine-tune did not target, and ERNIE-Image's card already has them. What was measured:

Held-out validation loss, a fixed 100-image panel, same noise and timesteps every time, scored during the fine-tune (step 1 of How it was made, before the speed transplant):

The safety edit's effect on quality — 124 adult prompts, the same prompt and seed on this model and on the same weights without the safety edit, each metric as a ratio (edited ÷ unedited). The bar was 0.98 for the first three and 0.90 for the face rate:

The edit cost nothing much on adult content; sharpness and face rate came out slightly higher.

Safety

Training data. Only commercial studio and studio-aggregator content was collected. Images a VLM tagger flagged as a possible minor were removed unless the flag traced to studio branding or a watermark on the image; the kept ones were re-checked with a calibrated age judge in a 5,535-image audit. All 299,996 training images went through Cloudflare's CSAM Scanning Tool (hashing) with zero matches. No prompt, caption or test in the project ever combined a minor with nudity or sexual content.

The weights. Before release, the model was trained to turn a request for a child into an adult: 1,200 SFW child prompts, most paired with an adult version of the same scene, fitted as a LoRA and merged into the weights. On 200 held-out child prompts, the share of images with a face read as under 18 fell from 80.5% to 6.0%, and 14 or under from 61.0% to 2.0%. Adult prompts lost nothing measurable (see Evaluation). The edit makes the model far less likely to draw a child; it does not make it impossible.

Quantized files. Four of them were measured the same way: 8.5% to 11.0% under 18 on the same 200 prompts, against 6.0% for the release measurement, none significantly different in a paired test. Every other official file has weights at least as close to the original as one of those four. On the 124 adult prompts, NVFP4 gave 2 frames with a face read as under 18 (the release gave 0), one of them from a prompt with no person in it. Per-file results: SAFETY.md.

Not covered. No prompt-attack search or relearning test was run on these weights. The edit increases the cost of misuse; it does not make it impossible, and the license forbids trying.

The method, the data audit, a likeness probe and the bugs we found in our own measurements are in SAFETY.md. To report misuse or a safety problem with the weights, please email abuse@verboa.io.

Limitations and biases

• Realism over range. This is a photographic model: the adult set is photographs only, and the general set's 8,757 artworks (WikiArt, The Met) are a small share. ERNIE-Image itself is better for drawn styles.
• Men are a smaller share of the data (16,253 images from two dedicated male sources; male-only images were repeated 4× in training). Male anatomy is slightly less reliable than female. Check male images closely.
• Partnered scenes drift. The model can substitute a neighboring act for the one named, add a third person to a scene when asked for as a couple, or give two people of the same sex the same face. Several seeds may help.
• Faint signatures. Although watermarked training images were cropped or dropped, the model occasionally draws a faint handwritten-style signature. Check corners before publishing an image.
• Unmarked defaults. The adult set's subjects, as the tagger labeled them, are 83.2% white, 10% Hispanic, 3% Asian and 2% Black; the model's own mix with no ethnicity in the prompt was not measured.
• Public figures: partial likeness for the most photographed people, inherited from ERNIE-Image; see Likeness.
• No prompt enhancer means short, under-specified scene prompts are weaker than in ERNIE-Image's default pipeline. Please describe the scene for accuracy.
• Literal readings. Slang is understood in context, but a bare word can be read literally ("chick" has produced a chicken before). Add a clarifying word.
• Text, hands, counting: good but not perfect; explicit scenes with three or more people lose anatomical coherence more often.
• Resolution: trained at about 1 MP. Larger sizes need a second pass or an upscaler.
• No trainable base in this release. This is a few-step model; LoRAs trained on it can behave oddly. A base model meant for fine-tuning comes with Verboa Image 2.0.

License

The weights are released under the CreativeML Open RAIL++-M License with one additional use restriction (LICENSE.md). Every word of the standard license is unchanged except that Verboa AI added a first item to its use restrictions (Attachment A): no sexually explicit, nude or intimate imagery of a real, identifiable person without that person's explicit prior consent. Because of that addition this is not the unmodified OpenRAIL++-M, and anyone who redistributes the model or builds a derivative must pass on this version, including the added restriction.

Read LICENSE.md for the terms; the use restrictions in Attachment A are the part that travels with every copy and derivative.

Upstream components and their licenses: ERNIE-Image and ERNIE-Image-Turbo transformer weights and pipeline (Baidu, Apache 2.0), Ministral 3 3B (Mistral AI, Apache 2.0), FLUX.2 VAE (Black Forest Labs, Apache 2.0). General-set images: fine-t2i curated (Apache-2.0 packaging of Unsplash, Pexels and Pixabay photos), CommonCatalog CC-BY, WebSight and DOCCI (CC-BY, sources linked under Training), The Met open access (CC0), and WikiArt, whose dataset states no license ("Data files © Original Authors"). See NOTICE.md for the Apache-2.0 components and what was modified. The license does not grant trademark rights.

Acknowledgments

Baidu's ERNIE-Image team for an Apache-licensed 8B DiT that fine-tunes like a wonder, and for a Turbo variant whose delta transplants; Mistral AI and Black Forest Labs for the text encoder and VAE; the diffusers and ComfyUI maintainers for native support; the authors of Concept Ablation, ESD, Ring-A-Bell and UnlearnDiffAtk, whose work the safety pass builds on.

Citation

@misc{verboa-image-1.0,
  title  = {Verboa Image 1.0: a photorealistic 8-step adult model built on ERNIE-Image},
  author = {{Verboa AI}},
  year   = {2026},
  url    = {https://huggingface.co/verboa/Verboa-Image-1.0}
}

Support

There is currently no API or subscription behind this model. verboa.io has the official links and a Bitcoin donation address. We appreciate any and all donations, they directly fund our future training runs to bring the community better models. Issues and questions: GitHub issues, or contact@verboa.io. Misuse or safety problems: abuse@verboa.io.

Changelog

• 1.0: first release. Fine-tune EMA at update 11,338, ERNIE-Image-Turbo delta at 0.85, safety edit v6 (2,400 steps) merged. Files: see Downloads.