Introduction: The AI That Puts Holes in Clothes
In the previous two articles, we explored how illustration-based models have a persistent tendency to expose skin, and tested countermeasures using prompt control and img2img with Denoising strength adjustments.
This article picks up where those left off. The target: Illustrious-based models. The problem: rear-view compositions.
All we wanted was a sailor uniform with a looking-back pose. Instead, the model cut a hole in the back of the uniform — a hole that doesn’t exist in the garment’s actual design — and exposed skin through it. When we patched the hole with prompts, it loosened the neckline. When we locked the neckline, it painted a purple butterfly on the character’s back — completely unprompted.
After generating 35 test images and confronting this problem head-on, we finally achieved full suppression using ControlNet (Canny). Here’s the complete record.
Three Problems in Illustrious-Based Models
Let’s start by organizing the issues observed during testing. The following three problems occur frequently with Illustrious-based models (mikanIllustrious / waiIllustrious):
Ghost characters — Unrequested figures appear in the background, facing away from the camera. Even in explicitly solo compositions, a second person showed up in 3 out of 4 images in some batches.
Clothing destruction — Garments that should be intact become torn, loosened, or expose the back. This is a continuation of the phenomenon reported in the previous articles.
Pose and composition drift — The generated pose deviates from what was specified in the prompt.
This round of testing focused primarily on the first two issues: ghost characters and clothing destruction.
Test 1: Does Raising CFG Scale Fix It?
The first hypothesis tested was straightforward: “Raising CFG Scale increases prompt adherence, so the problems should decrease.”
Test Conditions
- Model: Illustrious-based (mikanIllustrious / waiIllustrious)
- Clothing: Sailor uniform (closed-type garment)
- Composition: Upper body, front-facing
- Body tags: Large bust and glamorous body type specified (intentionally increasing undressing pressure as a stress test)
- Solo enforcement:
(solo:1.3), (1girl:1.3)in positive prompt
CFG was tested at three levels — 7 (baseline), 8, and 9 — with 5 images each, for a total of 15 images under identical prompts.
Results
| CFG | Ghost | Clothing loss | Rear-facing | Quality loss | Total issues |
|---|---|---|---|---|---|
| 7 | 0/5 | 0/5 | 0/5 | 0/5 | 0 |
| 8 | 0/5 | 2/5 | 2/5 | 0/5 | 4 |
| 9 | 0/5 | 1/5 | 2/5 | 0/5 | 3 |
The hypothesis was overturned. Raising CFG made things worse.
CFG 7 was the only level with zero issues across all categories. At CFG 8 and above, rear-facing poses appeared, and clothing destruction followed.
This appears to happen because Illustrious’s training data likely contains a strong association between sailor uniforms and looking-back poses. Raising CFG amplifies adherence to that learned pattern, causing the model to decide “if she’s wearing a sailor uniform, she should be looking back.”
CFG doesn’t uniformly boost adherence to all tags. It tends to preferentially amplify combinations the model has learned strongly. This is worth keeping in mind.
On the positive side, ghost characters appeared in 0 out of 15 images. The (solo:1.3), (1girl:1.3) enforcement combined with negative prompts held firm.
Test 2: Dressing the Back — Fighting with Prompts
The CFG test confirmed that front-facing compositions keep clothing stable. The problem is the back.
Looking-back poses are essential for storytelling, so dropping them isn’t an option. The only path forward was to find a way to keep clothes intact on rear-view compositions in Illustrious.
Facing Reality
After removing looking back from the negative prompt and adding it to the positive, we generated 5 images.
Result: 4 out of 5 images had a hole cut into the back of the sailor uniform, exposing skin.
Holes — in a garment that has no such opening. The model invented a structural breach in the clothing to achieve exposure. The same pattern as the previous article where it put holes in a shirt, but more aggressive this time.
Test 3: Clothing Defense Module — Block One Exit, It Finds Another
At this point, we re-examined the core problem.
The issue isn’t “rear-facing poses” per se. It’s that closed-type garments combined with rear-view compositions trigger clothing destruction. A swimsuit in the same pose is fine — the model doesn’t break it because there’s no exposure gap. Swimsuits already show skin, so no gap exists. Sailor uniforms cover the back, creating a large gap, and the model’s undressing pressure fires accordingly.
Based on this understanding, we designed a “clothing defense module” — a set of additional prompt tags to deploy only when using closed garments in rear-view compositions.
Clothing Defense Module
Added to positive prompt:
(clothes intact:1.3), (fully dressed:1.3), (back covered:1.3), (no skin exposed on back:1.2),
Added to negative prompt:
(bare back:1.4), (backless:1.4), (back cutout:1.3), (open back:1.3), (exposed back:1.3), (torn clothes:1.3), (ripped clothes:1.3), (hole in clothes:1.3),
Results
| Item | No defense | With defense |
|---|---|---|
| Holes in clothing | 4/5 | 0/5 |
Hole creation was completely blocked. So far, so good.
The Model’s Workarounds
However, once its primary exit was sealed, the model found two alternatives.
Loosened neckline (5/5) — If the back won’t open, try the front. The area below the sailor collar was pulled open to expose skin.
Painted a butterfly on the back (2/5) — If skin can’t be shown, draw on it. A purple butterfly appeared on the character’s back, completely unprompted. It even color-matched the character’s indigo eyes.
The previous article reported “block the shoulder, it strips the male character.” This time: “block the hole, it paints a butterfly.” The whack-a-mole pattern — seal one exit and the model finds another — remains alive and well. Prompt-only control has clear limits.
Test 4: ControlNet — A Weapon from a Different Dimension
Prompts are requests. No matter how many requests you stack, the internal pressure toward skin-colored pixels doesn’t go away.
ControlNet operates not as a request but as physical enforcement.
Using Canny (edge detection), the clothing’s contour lines are burned into the image structure. The shape of the sailor collar, the sleeve outlines, the skirt boundaries — all become hard lines that the model must follow during rendering. Painting skin where fabric exists becomes structurally difficult.
Settings
- ControlNet Unit 0: Canny (locks clothing structure)
- ControlNet Unit 1: OpenPose (locks pose)
- Reference image: An image where the clothing defense module successfully maintained the garment
The key parameter is Canny’s Control Weight. Too low and the model ignores it; too high and the output becomes a copy of the reference.
Results
| Item | Canny Weight 0.4 | Canny Weight 0.5 |
|---|---|---|
| Clothing destruction | 0/5 | 0/5 |
| Neckline loosening | Partial | 0/5 |
| Foreign objects (butterfly etc.) | 0/5 | 0/5 |
| Image quality | Good | Good |
At Canny Weight 0.5, every problem disappeared.
Zero holes. Zero neckline loosening. Zero butterflies. Zero quality degradation. All 10 images achieved a fully dressed sailor uniform in a looking-back composition.
The Decision Table: When to Use What
The final conclusion is straightforward.
| Composition | Garment type | Required countermeasure |
|---|---|---|
| Front-facing | Any | Prompts only (no ControlNet needed) |
| Rear-facing | Swimwear / exposed clothing | Prompts only (no ControlNet needed) |
| Rear-facing | Uniforms / closed-type clothing | ControlNet required (Canny 0.5 + OpenPose) |
ControlNet is only necessary for one specific scenario: closed-type garments in rear-view compositions.
ControlNet requires a reference image, but once you secure one good image, you can reuse it for all subsequent generations. The practical workflow is to stock one good rear-view reference per garment — and free yourself from the generation lottery.
Confirmed Parameters
| Parameter | Value |
|---|---|
| CFG Scale | 7 |
| Sampling method | DPM++ 2M |
| Schedule type | Karras |
| Steps | 25 |
| ControlNet Canny Weight | 0.5 (closed garments × rear-view only) |
| ControlNet Canny End Step | 0.8 |
| ControlNet OpenPose Weight | 0.7–0.8 |
Key Findings from This Test
CFG Is Not a Uniform Amplifier
The intuition that raising CFG improves prompt adherence across the board didn’t hold in this test. CFG can amplify strongly-learned garment × pose associations within the model. For Illustrious-based models with sailor uniforms, CFG 7 was the most stable setting.
Undressing Pressure Is Gap-Driven
The larger the gap between the model’s expected exposure level and the actual exposure level, the more likely it is to break the garment. Swimsuits stay stable even in rear-view because there’s no gap. Sailor uniforms get holes because the gap is large.
Prompt-Based Countermeasures Have Limits
Seal one exit and the model finds another. The previous article saw it strip a male character; this time it painted a butterfly. Prompt-only whack-a-mole has no end.
ControlNet Is a Structural Solution
Where prompts are requests, ControlNet (Canny) physically enforces clothing contour lines. Even under the harshest conditions — closed garments in rear-view — Canny Weight 0.5 achieved complete suppression.
Closing
Across this article and the two before it, quite a bit of practical knowledge has accumulated in “the battle to make AI keep its clothes on.”
Front-facing control with prompts. Rear-facing structure locked with ControlNet. Fine-tuning with img2img and Denoising when needed. The more tools in hand, the wider the range of options.
After 35 test images, the conclusion was a decision table that switches the countermeasure level based on the combination of composition and garment type. There’s no need to apply the same heavy solution to everything — the key is being able to say “ControlNet only where it’s needed.”
Stock one reference image per garment, and subsequent generations become much more stable. If you’re struggling with clothing destruction in Illustrious-based models, start with the clothing defense module prompts and work up from there.
