Edge Cases

Updated: 2026-05

1. About This Page

The third installment in the “Fun Experiment” trilogy. Edgy — A collection of experiments that intentionally aim to “break” or “cause failure.” While similar to the previous page (Algorithm Exposure), this one targets “beautiful failure” resulting from the inappropriate use of ControlNet, LoRA, and workflow structures.

In class, students engage in a form of flipped learning, discovering that “by learning how to make AI uncontrollable, they can actually gain insight into how it works.”

2. Topic A: Conflicts Between ControlNets

We discussed how to stack multiple ControlNet models. We intentionally create conflicting combinations.

Instructions

• ControlNet 1: “Standing pose” from OpenPose
• ControlNet 2: “Sitting pose” from OpenPose
• Apply both with a strength of 1.0
• Result: The AI attempts to satisfy both conditions, resulting in a chimera pose where the character appears to be sitting while standing

An experiment to observe how AI fails when faced with conflicting constraints.

3. Example B: Blasting ControlNet with a strength of 5

The strength of ControlNet nodes is typically between 0.7 and 1.0. Set this to 3 or 5.

• Result: The edges and colors of the control image are directly imprinted on the output
• The result of the noise and control image being forcefully blended
• An artistically interesting noise and glitch effect

4. Idea C: Contradictory Prompt

Intentionally include conflicting instructions in the prompt.

Examples:

• a smiling angry warrior, both happy and sad, with eyes that are both closed and wide open
• a square circle, both red and blue at the same time
• a cat that is also a dog, rendered in photorealistic detail

Since AI tries hard to satisfy both, it results in hybrid organisms and color confusion.

5. Prompt D: Write the main prompt in a negative tone

An experiment to swap positive and negative.

• Positive: (blank) or image
• Negative: a cat

What will an AI that’s been told “Don’t draw a cat” draw? Something abstract—or an eerie “absence of a cat.”

6. Example E: Repeating the same prompt 100 times

a cat a cat a cat a cat a cat a cat a cat a cat a cat a cat a cat a cat ...

Repeat the same word until the CLIP token length limit (77 tokens by default, up to 225 tokens) is reached.

• Result: Over-saturation or strange emphasis occurs
• An extreme version of the “emphasize” prompt technique (cat:1.5)

7. Challenge F: Causing img2img to Enter an Infinite Loop

This is similar to the “Algorithm Exposure” J method, but in this case, the denoise value is increased with each iteration.

• 1st iteration: Original image → img2img (denoise 0.3)
• 2nd iteration: Output → img2img (denoise 0.5)
• 3rd iteration: Output → img2img (denoise 0.7)
• 4th iteration: Output → img2img (denoise 0.9)
• 5th iteration: Output → img2img (denoise 1.0, i.e., completely forget)

It is visualized as a series of images depicting the process of gradually forgetting the original image. It also stands on its own as a work of art.

8. Task G: Repeat VAE encoding and decoding 10 times

Compress the image into latent space using a VAE → Decode it back into an image using the VAE. Repeat this 10 times.

• 1st time: Almost identical
• 3rd time: Details begin to fade
• 5th time: Colors begin to change
• 10th time: A “VAE-averaged visual style” that bears little resemblance to the original image

A visualization demonstrating that VAE is not a perfectly reversible transformation. Visually understanding that “compression equals degradation.”

9. Tip H: Make Full Use of LoRA

Stack multiple LoRAs with a “strength of 1.0” for all of them (typically, use 2 to 3 LoRAs with a strength of 0.5 to 0.8).

• LoRA: Studio Ghibli + Cyberpunk + ’80s Movies + Pokémon Pixel Art + Enhanced Details
• All set to 1.0
• Result: The styles clash, resulting in utter chaos

You can see how the AI handles style conflicts between LoRAs.

10. Idea I: A Triple Combination of ControlNet, img2img, and LoRA

Combine all controls to their maximum settings.

• ControlNet: OpenPose strength 1.0
• img2img: Denoise from a different image 0.6
• LoRA: Strong style LoRA strength 1.0
• Prompt: A different subject

Give the AI four conflicting instructions at the same time. The output is unpredictable, but this is the moment when you can see the limits of the AI system.

11. Tip J: Reduce the resolution to 64×64

The standard resolution for SD 1.5 is typically 512×512. This is reduced to 64×64 or 128×128.

• Result: The number of pixels is too low, resulting in an image that is mostly noise or resembles a very small thumbnail
• Issues arising from deviating significantly from the resolution used during training

Conversely, if you increase the resolution to 4096×4096, the model either fails to run due to a memory shortage error or produces a repeating pattern (high resolutions are outside the training scope of the base model).

12. Scenario K: Comparing all models with a fixed seed of 0

Seed 0 is a common value in the AI world. There are certain images that tend to appear more frequently.

• The same prompt is used for all models (SD 1.5, SDXL, Flux dev, Z Image Turbo) with seed 0
• By comparing them, you can see each model’s “preferences at seed 0”

13. The Educational Significance of This Page

Edge computing goes beyond “the knowledge needed to create beautiful works” and is about “understanding the limits of AI systems.”

• How effective is it?
• Where does it break down?
• How should we handle contradictions?
• What are the limitations of the training data?

Once you’ve gained this hands-on experience, you’ll be able to calmly identify the cause when commercial tools like Runway “don’t work properly” or “produce strange results.”

14. Credit Budget

Even for edgy content, SD 1.5 is plenty. You can bring about five segments into class, with each segment costing 2 to 5 credits.

There’s no need to have each student try them all. The teacher should find one memorable example in advance and demonstrate it during class.

15. What’s Next

• To Runway — Apply the skills you’ve learned in Comfy Cloud to the Runway exercises