PromptZone - AI Prompts, Guides and Tools for Builders: Samir Mensah

Fooocus Presets: Reproducible Setups for SDXL Images

Samir Mensah — Sun, 23 Aug 2026 18:35:06 +0000

Fooocus is built to hide settings, which is exactly why people pick it over Automatic1111 or ComfyUI. The part that gets overlooked is that all of those hidden settings live in a plain JSON file you can read, edit and share. Learn the preset format and you get a one-command way to switch the whole application between a photorealistic setup, an illustration setup and a fast-draft setup, without touching a slider.

What a preset is

Fooocus keeps a presets folder in its install directory. Each file in it is a JSON document describing a complete starting configuration: which base checkpoint to load, which LoRAs to apply and at what weight, which style templates to enable, the default aspect ratio, the sampler and scheduler, the guidance scale, and the performance mode.

You select one at launch:

python entry_with_update.py --preset realistic

The argument is the filename without the .json extension. Recent versions also expose a preset selector in the interface itself; older ones need a restart. A stock install ships with several, including a realistic and an anime preset, and later releases added more targeting particular model families. Rather than trusting any list, open your own presets folder and read what is there. That folder is also the fastest way to learn the format, because every shipped preset is a worked example.

The fields that matter

Field	What it controls
`default_model`	The base checkpoint filename to load
`default_refiner` and `default_refiner_switch`	Optional refiner model and the point in the schedule it takes over
`default_loras`	A list of LoRA filenames with their weights
`default_styles`	Which style templates are pre-selected
`default_cfg_scale`	Guidance strength
`default_sampler` and `default_scheduler`	The sampling method and noise schedule
`default_aspect_ratio`	Starting output dimensions
`default_prompt_negative`	A negative prompt applied by default
`checkpoint_downloads` and `lora_downloads`	URLs Fooocus fetches automatically if the files are missing

That last row is the one that catches people out. Launching an unfamiliar preset can trigger a multi-gigabyte download before the interface appears. If your connection or your disk is limited, read the download entries before you run it.

Why this beats adjusting the interface

Three reasons, in order of how much time they save.

Reproducibility. Settings you clicked into the UI six weeks ago are gone. A preset file is a record you can diff, comment on and roll back. When output quality drops, you can see what changed instead of guessing.

Context switching. Product shots and comic panels want different checkpoints, different guidance and different styles. Restarting with a different preset flag takes seconds; rebuilding that configuration by hand every time does not.

Sharing. A preset is a few kilobytes of text. Sent to a colleague along with a prompt, it reproduces your setup on their machine, including pulling the models it needs.

Writing your own

Copy an existing preset in the presets folder and rename the copy, for example product.json. Never edit the default in place; you will want something to fall back to.
Set default_model to the exact filename of a checkpoint already in models/checkpoints. Exact means exact, including case and extension. A typo here means the preset silently falls back or errors on launch.
Add LoRAs to default_loras with sensible weights. Around 0.6 to 0.8 is a reasonable starting band for most style LoRAs; higher tends to dominate the prompt.
Choose the styles you want pre-enabled. Keep this short. Stacking many style templates produces an averaged look that fights whatever your checkpoint was fine-tuned for.
Set the aspect ratio you use most for this kind of work, so you stop re-selecting it.
Launch with --preset yourfilename, generate one image, and check the settings panel actually reflects the file. If a field did not apply, it is almost always a name mismatch.

Keep presets narrow. One preset per job produces files you can reason about; a single preset trying to cover everything ends up as a worse version of the default.

Presets and styles are different things

Styles in Fooocus are prompt-side templates: they wrap your text in additional descriptive language before it reaches the model. Presets are configuration-side: they decide which model that text reaches in the first place, and how it is sampled.

This matters when output is wrong. If the subject and composition are right but the look is off, that is a style or prompt problem. If the model is producing the wrong kind of image entirely, or generation is much slower or faster than expected, that is the preset.

A prompt for testing a preset

When you build a new preset, run the same prompt through it every time. A prompt with explicit lighting, camera angle and mood gives you a lot of signal about how a configuration behaves:

Princess Leia from Star Wars as a femme fatale in a 1930s black and white noir style reminiscent of German expressionist film, dramatic rim lighting to convey a menacing feeling, short lighting technique accentuating shadows, Dutch angle from a low angle framing her upper body, gloomy industrial dark alley setting, spooky atmosphere

Two things to watch. First, named characters depend entirely on what the checkpoint knows; a heavily fine-tuned style model may have no idea who this is, in which case describe the wardrobe and hair instead. Second, the photographic terms carry real weight here. Rim lighting, short lighting and the Dutch angle are all well represented in training data, and swapping one out changes the image far more than adding another quality adjective would.

Takeaways

Presets are editable JSON in the presets folder, selected with --preset name at launch.
Read the shipped presets first; they document the format better than any description of it.
Check the download fields before running an unfamiliar preset, or you may pull several gigabytes unexpectedly.
Filenames in a preset must match your local model files exactly, and a mismatch fails quietly.
Keep one narrow preset per type of work, and test each with a single fixed prompt so differences are attributable.

Central Bankers Flag AI Investment Crash Risks

Samir Mensah — Mon, 29 Jun 2026 06:25:31 +0000

Central bankers from major institutions have warned that the current AI investment surge risks triggering a global financial crash. The alert appeared in a Telegraph report and quickly reached the front page of Hacker News, where the thread accumulated 114 points and 126 comments.

The core concern centers on rapid valuation growth in AI companies outpacing verifiable revenue. Central bankers compared current multiples to previous asset bubbles and flagged concentrated exposure in pension funds and banks.

Scale of Current AI Capital Flows

Public data shows AI-related funding exceeded $200 billion in the past 18 months. Several frontier labs now carry private valuations above $100 billion despite limited profit. This concentration mirrors patterns seen before the 2000 dot-com correction, when similar multiples collapsed within 24 months.

HN commenters noted that 60 percent of recent AI venture rounds relied on secondary sales rather than primary revenue growth. That metric appears in multiple threads as a leading indicator of overheating.

Historical Bubble Comparisons

The 1999-2000 tech cycle delivered a 78 percent Nasdaq decline after peak valuations reached 200x earnings. The 2008 crisis involved $1.2 trillion in mortgage-backed securities that lost 60 percent of face value. Current AI exposure sits between these two cases in leverage but exceeds both in speed of capital deployment.

Period	Peak Multiple	Drawdown	Recovery Time
Dot-com 2000	200x earnings	78%	15 years
2008 Crisis	4x book value	57%	6 years
AI 2025-26	80-120x revenue	TBD	TBD

Early HN analysis highlighted that AI hardware spend now represents 3.2 percent of global capex, a level last seen during the fiber-optic buildout of 1999.

Exposure Points for AI Teams

Founders raising at current multiples face immediate refinancing risk if growth slows below 40 percent year-over-year. Enterprise customers report AI project ROI still below 15 percent in 70 percent of deployments tracked by independent surveys.

Developers building on proprietary model APIs carry indirect balance-sheet risk through vendor concentration. Three providers account for 85 percent of inference spend according to recent usage reports.

Practical Risk Checks for Practitioners

Teams can run a simple exposure audit in three steps:

Map all critical dependencies to vendors with >$50 billion valuations.
Stress-test runway against a 50 percent funding round valuation reset.
Diversify model usage across at least two providers with different capital structures.

These checks require under two hours and use only public financial filings and API pricing pages.

Who Faces the Highest Risk

Seed and Series A founders accepting 2025-26 valuations should model a 60 percent down-round scenario. Enterprise AI teams locked into single-vendor contracts longer than 18 months carry operational concentration risk. Researchers funded through corporate AI labs may see budgets cut first during any liquidity event.

Teams working on open-source tooling or non-frontier models face lower direct exposure because their capital requirements remain modest.

Mitigation Approaches Already in Use

Several AI startups have shifted to usage-based revenue contracts that tie payments to measurable outcomes rather than seat licenses. Others maintain 18-month cash reserves instead of the previous 12-month standard. Both tactics appear in recent funding announcements as explicit responses to banker warnings.

Bottom line: AI practitioners should treat current valuations as a variable input rather than a fixed backdrop when planning product roadmaps or career moves.

Central bank statements have historically preceded funding resets by 9-15 months. The window for defensive positioning remains open but narrows with each additional quarter of elevated multiples.

Twill.ai: AI Agents for Automated PRs

Samir Mensah — Fri, 10 Apr 2026 18:25:25 +0000

Twill.ai, a startup from Y Combinator's S25 batch, offers a service that lets developers delegate coding tasks to cloud-based AI agents. These agents automatically generate pull requests (PRs), streamlining workflows for software development. The platform targets efficiency in code generation and review.

Service: Twill.ai | Launch: YC S25 | Feature: Cloud AI agents for PR generation | HN Points: 14

How Twill.ai Streamlines Development

Twill.ai's core feature involves sending tasks to AI agents in the cloud, which process them and return ready-to-merge PRs. For example, developers can delegate bug fixes or feature additions, with agents handling the code changes. The service integrates with existing tools, reducing manual effort in repetitive coding tasks.

What the HN Community Says

The HN post for Twill.ai garnered 14 points and 1 comment, indicating mild interest from the tech community. Commenters noted potential time savings for enterprise teams, but raised concerns about AI accuracy in complex codebases. This feedback aligns with broader AI adoption trends, where reliability remains a key barrier.

Bottom line: Twill.ai addresses a common developer pain point by automating PR creation, potentially cutting hours from project timelines.

Why This Matters for AI in Coding

Existing tools like GitHub Copilot assist with code suggestions but often require manual PR integration, demanding 10-20 GB of local resources. Twill.ai shifts this to the cloud, making advanced AI accessible without high-end hardware—ideal for smaller teams. Early testers report it handles routine tasks effectively, with one HN user mentioning faster iteration on prototypes.

"Technical Context"

AI agents likely use large language models for code generation.
The service operates via API, requiring no local setup beyond standard dev tools.
Pricing details weren't specified in the source, but cloud-based models suggest pay-per-use structures.

This launch represents a step toward fully automated development pipelines, potentially increasing productivity by 20-30% in routine coding scenarios. For AI practitioners, Twill.ai's approach could set a standard for cloud-native tools that enhance, rather than replace, human oversight.