Fable 5.1 Generates Minecraft Shaders: Stunning Results but Max Mode Is a Trap

Using AI to Create Minecraft Shaders: A Bold Experiment

Using large language models for game shader development has become a hot topic in the Minecraft modding community. Traditionally, building a Minecraft shader pack requires deep knowledge of OpenGL rendering pipelines, GLSL shader languages, and various lighting algorithms — a fairly high barrier to entry. In this test, a Bilibili content creator handed that entire job to Fable 5.1, asking it to generate a complete shader pack for OptiFine (a popular shader mod) from scratch.

The setup was straightforward: using the creator's own API relay and key, they issued a direct prompt to the model and watched to see whether it could independently produce a working shader solution with no existing shader code as reference. The result, in the creator's own words: "One round of conversation later, it looked good enough to use as a wallpaper."

Pitfall Report: The Efficiency Trap of Max Thinking Mode

Before evaluating the final output, the creator spent considerable time venting about the experience — a section that's highly valuable for anyone looking to try similar AI-assisted shader generation.

Max Mode: Two Hours, Zero Output

The core lesson: do not use Max thinking mode. The creator ran the model in Max mode for a full two hours and got nothing usable. Logs showed the model performed nearly 40+ web searches during that time, including 11 searches within a single hour.

This exposed a classic problem with current AI agents handling long-chain professional tasks: over-retrieval and runaway reasoning. The model got stuck in a loop of searching for information, verifying it, then searching again — burning enormous compute without ever converging on a deliverable.

Token Cost: Nearly 20 Million Across Two Runs

The efficiency problem translated directly into cost. The creator reported consuming nearly 20 million tokens across two runs. For a shader generation task, that's a staggering amount, and it makes Max mode completely impractical for this use case.

Ultimately, the creator abandoned Max mode, started a fresh conversation, and used a standard high-thinking intensity (described as "high thinking strength, not the Max tier") — and ironically, this downgrade produced faster, more usable results.

Results: Shader Quality Exceeded Expectations

What genuinely surprised the creator was the quality of the final render. The generated Minecraft shader pack covered multiple environments and weather conditions, each of which was tested individually.

Full Coverage Across Day/Night and Weather

The model didn't just handle basic daytime lighting. Testing revealed:

• Night scenes: Dedicated handling was in place; the creator exclaimed "this is actually impressive" upon seeing the nighttime visuals;
• Rainy weather: Targeted adjustments were made, and the rainy atmosphere came through convincingly;
• The Nether dimension: The model independently adapted to a summer/Nether configuration mid-generation, with other dimensions adjusted accordingly — though the changes there were less dramatic;
• Thunderstorms: This was the weakest area, with relatively minimal visual changes.

The fact that the model handled each environment state separately suggests it genuinely understood that a shader pack needs to cover different in-game scenarios — rather than simply applying a single universal filter.

Advanced Shader Features Like the "Darwin Effect"

The creator specifically checked for several advanced shader characteristics. The so-called "Darwin effect" (a community term for a particular type of light ray/depth-of-field behavior) was present in the output, and the rain effect implementation was satisfying. The level of detail prompted repeated remarks of "this is actually pretty good."

Shortcomings: Rough Edges Still Need Work

Impressive as it was, the AI-generated shader pack wasn't flawless:

• Water surfaces: Underwater visuals had obvious issues and would need further correction;
• Missing entities: The creator noted "my fish seem to have disappeared," indicating that rendering for certain in-game entities broke down;
• Thunderstorm weather: As mentioned, lighting changes during thunderstorms were insufficient.

Interestingly, all of this was produced in non-Max mode. The creator speculated that Max mode — if it could actually converge — might yield even better results. But given Max mode's current efficiency problems, that hypothesis remains impossible to verify.

Practical Implications and Recommendations for AI-Assisted Shader Development

This experiment sends a fairly positive signal: even without any existing shader code as a reference, Fable 5.1 can independently generate a working Minecraft shader pack in a single conversation — one that covers day/night cycles, rain, multiple dimensions, and advanced features like the Darwin effect. In the creator's words, the final product "is at least usable" and could even serve directly as a wallpaper.

At the same time, the practical pitfalls are clear: your choice of thinking mode can make or break the whole attempt. Blindly chasing the highest-tier Max mode can lead to retrieval loops, two hours of zero output, and a massive token bill. For anyone looking to replicate this kind of experiment, choosing the right mode and managing costs is probably more important than simply throwing more compute at the problem.

Overall, AI is demonstrating increasingly capable out-of-the-box performance in specialized, graphics-intensive creative domains. While details like water surfaces and entity rendering still need human polish, the fact that "generate a shader pack from a single prompt" has moved from concept to something genuinely usable is a development worth paying attention to.