PromptZone - AI Prompts, Guides and Tools for Builders: Seojun Zhao

Is 30% of arXiv Now AI-Written?

Seojun Zhao — Mon, 20 Jul 2026 18:25:22 +0000

A recent analysis of arXiv submissions found that over 30% of new papers exhibit detectable AI writing patterns. The findings first appeared in a Hacker News thread that drew 100 points and 65 comments.

Scale of the Shift

The study examined recent uploads across multiple categories. Computer science and machine learning sections showed the highest rates, exceeding the overall 30% average. Earlier estimates from 2022 placed AI-assisted text below 10% in the same corpus.

Detection Method

The approach relies on statistical markers such as token probability distributions and sentence-level perplexity scores. These signals flag text generated by current large language models without requiring access to the original model weights. The method was calibrated against known human-written and AI-generated samples from prior years.

Key Numbers Reported

30%+ of new submissions flagged overall
Highest concentration in cs.AI and cs.LG categories
65 comments on the Hacker News thread discussing methodology
100 points accumulated within the first day of posting

Category	Estimated AI Rate	Change Since 2022
cs.AI	38%	+28 points
cs.LG	35%	+25 points
Physics	22%	+14 points
Mathematics	18%	+11 points

Detection Tools Compared

Several public detectors exist for checking similar patterns. ZeroGPT and GPTZero focus on probability curvature. Originality.ai adds citation and style consistency checks. The arXiv study method differs by operating on full-document statistics rather than sentence fragments.

Who Should Track This

Researchers preparing submissions to competitive venues need to review their drafting process. Journals and conferences evaluating policy on AI assistance can use the numbers to set thresholds. Reviewers spotting unusually uniform prose now have aggregate data to reference when raising questions.

Practical Next Steps

Run any draft through at least two independent detectors before submission. Keep version history that shows human edits and source notes. For borderline cases, add explicit statements on AI tool usage in the acknowledgments section.

Bottom line: arXiv now contains a measurable volume of AI-generated text that existing statistical tools can surface at scale.

The trend will force clearer disclosure rules and better detection benchmarks within the next submission cycle.

GPT-5.6 Edges Grok 4.5 in App Build-Off

Seojun Zhao — Sat, 11 Jul 2026 00:25:17 +0000

A recent Hacker News thread tracked GPT-5.6, Grok 4.5, Claude, and Muse Spark building the same four applications under identical prompts.

The exercise produced 130 points and 74 comments focused on measurable differences in code structure, error rates, and revision cycles.

Build-Off Task Breakdown

Each model received the same four specifications: a task manager with real-time sync, a minimal analytics dashboard, a file-upload API with validation, and a lightweight chat interface with persistence.

Prompts stayed fixed across runs. No model-specific tuning occurred.

Output Metrics from the Thread

Participants logged concrete results across 12 total model runs.

GPT-5.6 completed all four apps with the fewest follow-up prompts (average 1.8 revisions).
Grok 4.5 produced the longest initial code blocks but required 3.4 revisions on average.
Claude delivered the cleanest TypeScript types in two of the four tasks.
Muse Spark showed the fastest first-token response but the highest rate of incomplete functions.

How to Replicate the Test

Clone the four app specs from the original thread. Feed each model the exact prompt sequence while recording revision count and final test coverage.

Run the same unit-test suite on every output. Track tokens used and wall-clock time per iteration.

Observed Tradeoffs

GPT-5.6 produced the most consistent folder structures and fewer runtime errors. Grok 4.5 generated richer feature sets on the first pass but introduced more merge conflicts during integration.

Claude excelled at strict type safety yet sometimes over-engineered simple endpoints. Muse Spark stayed fastest for prototypes but left more TODO comments.

Who Should Run Similar Tests

Teams selecting a primary coding model benefit from repeating the four-app exercise on their own stack. Solo developers already satisfied with one provider can skip the overhead.

Organizations evaluating cost per successful deployment should weight revision count more heavily than raw generation speed.

Model Selection Verdict

The thread data indicates GPT-5.6 currently leads on end-to-end reliability for small-to-medium internal tools, while Grok 4.5 remains competitive when maximum feature density on the first attempt matters more than polish.

Bottom line: Re-running the exact four-app prompts on current frontier models gives developers the clearest signal for production use.

Early comments note that prompt length and example count affected outcomes more than model size alone.

Pulpie Delivers Pareto-Optimal Web Cleaning Models

Seojun Zhao — Tue, 07 Jul 2026 00:25:22 +0000

Pulpie models surfaced on Hacker News with an 81-point Show HN thread and 19 comments. The release focuses on Pareto-optimal classifiers that remove low-quality or toxic content from web-scale datasets while preserving usable text volume.

What It Is and How It Works

Pulpie trains multiple models across quality, toxicity, and duplication axes. Each model outputs scores that let users select operating points on the Pareto front rather than a single fixed threshold.

The approach trains lightweight classifiers on curated subsets, then evaluates trade-offs between retained tokens and contamination rates. Users apply the models sequentially or in ensemble during Common Crawl processing.

Benchmarks and Key Numbers

Early results show retention rates between 38% and 72% of raw tokens depending on the chosen front point. Toxicity flagging reaches 94% recall at the strictest setting while keeping false-positive rates under 6%.

Metric	Strict Front	Balanced Front	Lenient Front
Token retention	38%	55%	72%
Toxicity recall	94%	87%	71%
Duplicate removal	82%	74%	61%
Model size (parameters)	340M	340M	340M

The thread notes these numbers come from a 100M-document subsample of recent Common Crawl.

How to Try It

The models are available via the project repository linked in the HN post. Users download weights, run inference with a provided Python script, and pipe scores into existing filtering pipelines.

Typical command flow:

python pulpie/score.py --input shards/ --output scores/ --front balanced

Integration requires under 50 lines of additional code for most Ray or Spark workflows.

Pros and Cons

Pros
- Explicit Pareto curves let teams choose exact quality-volume trade-offs.
- 340M parameter size runs on a single A100 in under 3 hours for 10B tokens.
- Open weights reduce reliance on proprietary filters.
Cons
- No built-in multilingual support beyond English and German.
- Requires separate handling of code and math content.
- Evaluation limited to one Common Crawl snapshot.

Alternatives and Comparisons

Pulpie differs from prior filters such as the original C4 quality classifier and the more recent FineWeb-edu pipeline.

Feature	Pulpie Pareto	C4 Classifier	FineWeb-edu
Multiple fronts	Yes	No	No
Toxicity + quality	Combined	Quality only	Quality only
Model size	340M	1.5B	1.5B
Open weights	Yes	Yes	Partial

Teams that need only English quality filtering may still prefer the lighter C4 baseline.

Who Should Use This

Research labs building custom pre-training corpora benefit most. Production teams already satisfied with existing toxicity APIs can skip it. Organizations needing strict regulatory compliance should validate Pulpie scores against their own red-team datasets first.

Bottom line: Pulpie gives practitioners controllable trade-offs instead of one-size-fits-all web cleaning.

The release marks a practical step toward reproducible, tunable data pipelines rather than opaque proprietary filters.

HiDream-I1 License Guide to Open Weights and Commercial Use

Seojun Zhao — Sun, 05 Apr 2026 18:25:18 +0000

HiDream-I1 publishes its code and transformer weights under MIT terms. Its full generation pipeline also uses separately licensed components, including Llama. Repository Model card

What are the key facts about HiDream-I1 licensing?

Field	Verified information
Developer	HiDream.ai. Repository
Released	April 7, 2025. Repository
Type	Text-to-image generation; Full plus distilled Dev and Fast variants. Repository
Size or parameters	17 billion parameters for the image model. Full model card
License and access	MIT for code and transformer weights; other pipeline components have separate terms. Downloads are available on Hugging Face. Full model card
Where it runs	Local Python inference and Gradio; a Diffusers integration is documented. Repository Diffusers documentation

For a deployment decision, record three things: the chosen checkpoint, the licenses shipped with its components, and the output checks required by your project.

What does the HiDream-I1 release let you use?

The Full model card describes photographic, cartoon, and artistic image generation, and publishes evaluations covering prompt following and visual preference. These are developer-reported results, useful for choosing test cases rather than guaranteeing the quality of your next image. Full model card

Record the exact Full, Dev, or Fast model identifier with each approved output. Include the dependency revisions so another contributor can identify the environment used.

For an illustration commission, test the actual objects and backgrounds you need. Check headline space, visual clarity, and whether the image communicates the intended action.

The HiDream-I1 repository license permits changes and redistribution subject to retaining the required notices. Keep those notices with any software or model package you distribute. License file

Which HiDream-I1 components have separate license terms?

The transformer license does not describe every dependency. HiDream's model card identifies a FLUX-derived VAE and T5 and Llama text encoders with their own terms. The repository also warns that automatic Llama downloads require accepting the model agreement and authenticating a Hugging Face account. Full model card Repository

Keep a component inventory when packaging an application: model identifier, source repository, revision, license file, and any required account access. This is a practical way to avoid discovering a missing dependency only when moving from a development machine to a deployment environment.

A parameter count also does not specify the memory requirement of a complete running pipeline. Diffusers exposes multiple text encoders in HiDreamImagePipeline, alongside the transformer and VAE. Hardware planning should therefore start with the chosen implementation and precision, then be checked on the intended machine. Diffusers documentation

Use the documented sampling schedule for the selected variant. Full, Dev, and Fast use 50, 28, and 16 steps respectively in the reference implementation. Repository

How do you evaluate HiDream-I1 for a commercial project?

Choose the model variant and deployment target before collecting license records. The sibling HiDream-I1 setup guide covers downloading and encoder access.

For the developer's repository workflow, prepare the required Python and CUDA environment and install its dependencies, including Flash Attention. Obtain the required Llama access before running the example. The following commands follow the repository's documented entry point; they download and run a model when executed in a prepared environment. Repository

git clone https://github.com/HiDream-ai/HiDream-I1
cd HiDream-I1
python -m pip install -r requirements.txt
python -m pip install -U flash-attn --no-build-isolation
python inference.py --model_type full

After the initial example succeeds, make a small evaluation folder with the prompt, selected model, settings, and resulting image for each trial. Record failures as well as approved outputs. A collection containing only favorites cannot tell another reviewer how often the workflow needed revision.

Use a brief such as “An editorial illustration of a brass watering can beside a clay flowerpot, pale background, space above for a headline.” First inspect the objects and their relationship. Then try a photographic interpretation of the same scene. Keep the content stable so that the review focuses on the change in visual treatment.

Place the proposed illustration beside its intended headline and view it at publication size. Record any manual corrections required before approval.

Keep the approved configuration with the deliverable. At minimum, note the checkpoint, dependency versions, prompt, output dimensions, and any postprocessing. The aim is to make the next revision understandable to someone who did not perform the original experiment.

How do HiDream-I1 and FLUX.1-schnell licenses compare?

FLUX.1-schnell is another downloadable text-to-image model, published by Black Forest Labs under Apache 2.0. Its model card documents a short sampling schedule and a Diffusers example; HiDream publishes Full and distilled releases with separate schedules. Compare the deployment requirements and output behavior of the specific variants you would actually use. [FLUX.1-schnell card][flux] Dev card Fast card

For a visual workflow, the ComfyUI pillar provides the broader interface context. Choose the model separately from the interface: the same project may benefit from a scripted evaluation and a visual workflow for later creative iteration.

What should you know before adopting HiDream-I1?

What license covers HiDream-I1 code and weights?

HiDream-I1 code and transformer weights use the MIT license. The VAE and text encoders have separate terms that must also be checked for a complete deployment. License Model card

Can I use HiDream-I1 images commercially?

HiDream-I1's model card permits commercial use of generated images under its stated conditions. It also assigns responsibility for model use to the user. Full model card

Do Full, Dev, and Fast use different licenses?

HiDream-I1 Full, Dev, and Fast each specify MIT terms for their transformer weights. Each model card also requires compliance with the separate component licenses. Full Dev Fast

Does the 17B parameter count include all deployment components?

HiDream-I1's published image-model size is 17B parameters. The Diffusers pipeline also loads a VAE and four text encoders, so plan memory for the complete configuration. Full card Pipeline

Sources

[flux]: https://huggingface.co/black-forest-labs/FLUX.1-schnell

AI Model Releases Timeline

Seedream Prompts: A Practical Guide to Scenes and Image Edits

Seojun Zhao — Fri, 03 Apr 2026 18:25:56 +0000

Write Seedream prompts with a subject, action, and setting, or specify an edit and what to preserve. This guide covers ByteDance Seed's hosted Seedream 4.0 and 4.5 models. Prompt guide, Product page

What are the key facts about Seedream 4.0 and 4.5 prompts?

Fact	Verified detail
Developer	ByteDance Seed. Announcement
Released	Seedream 4.0: September 9, 2025; exact 4.5 release date not published on its cited product page. Announcement, 4.5 page
Type	Text-to-image generation, editing, and reference-based creation. Prompt guide
Size or parameters	Not published on the cited product pages. 4.0 page, 4.5 page
License and access	Hosted access under provider terms; no open weights supplied. Product page, API tutorial
Where it runs	Provider infrastructure reached through web services or APIs. Product page, API tutorial

What should you include in a Seedream prompt?

BytePlus recommends coherent scene descriptions and clear editing instructions. Its guide covers new images, reference-based creation, and related image sets, so choose the task before writing the prompt. Prompt guide

For a scene, begin with the subject and what it is doing. Add the setting, then the visual properties you care about. This keeps the description readable and gives you concrete details to inspect in the result.

An original example is: a florist wrapping yellow tulips at a wooden counter, with a large window on the left and muted blue walls. Use an editorial photograph style with soft daylight.

Decide whether each detail is necessary. If the counter material matters to the brief, keep it. If you would accept any counter, leave room for variation and concentrate on the florist, flowers, and composition.

For an edit, identify the target and the requested change. State what should remain stable, such as the person's pose or the position of objects in the room. Prompt guide

A proposed edit to the florist scene is: change the tulips to white while preserving the wrapping paper, hands, counter, and viewpoint. Compare the result with the source before introducing a second change.

What can a Seedream prompt control?

The official prompt guide recommends concise, precise language instead of repeated ornamental wording. It does not provide a universal accuracy gain for a particular prompt formula. Prompt guide

Treat the examples here as starting briefs, not tested performance claims. The right amount of detail depends on what your image must communicate and how you will decide whether the output succeeds.

A prompt also operates within the selected API's capabilities. Set size and generation mode through documented controls; do not assume prose alone changes an unsupported request option. API tutorial

Keep exact wording for visible text in a separate source note. After generation, compare every word with that note instead of accepting a layout merely because its lettering looks plausible.

When a reference contains several similar objects, describe the intended target unambiguously. The official guide recommends identifying what should change and what should remain. Prompt guide

Avoid diagnosing every failure as a need for a longer prompt. First identify the missed requirement, then decide whether it needs clearer wording, a different reference, or a simpler creative brief.

How do you write and test Seedream prompts?

Start with a one-sentence brief and a short acceptance list. For the florist example, that list might include tulips being wrapped, a visible counter, and light arriving from the left.

Then choose the documented model and output settings in your hosted service. The Seedream API access guide covers ModelArk setup and the handling of returned images.

The following adapted request sends an original scene prompt to the documented Seedream 4.0 endpoint. Configure ARK_API_KEY according to BytePlus's tutorial before using it. API tutorial

curl --fail-with-body https://ark.ap-southeast.bytepluses.com/api/v3/images/generations \
  -H "Authorization: Bearer $ARK_API_KEY" \
  -H 'Content-Type: application/json' \
  -d '{
    "model": "seedream-4-0-250828",
    "prompt": "A florist wrapping yellow tulips at a wooden counter, large window on the left, muted blue walls, editorial photograph.",
    "size": "2K",
    "response_format": "url",
    "sequential_image_generation": "disabled"
  }'

Review the image in the same order as the brief: subject, action, setting, then appearance. Note the most important failure before adjusting the prompt.

If the flowers are present but nobody is wrapping them, clarify the action. If the action is clear but the window dominates the image, revise the framing. Keep unrelated parts of the prompt stable during each comparison.

For reference editing, provide the image through the documented image field and use a change instruction. For multiple references, assign each source a subject, setting, or style role. API tutorial, Prompt guide

Write a small reference map for yourself before submitting: subject from the first image, background from the second. Check that the written prompt expresses those roles without relying on your private notes.

For related outputs, BytePlus's guide recommends explicitly requesting a series or set. Configure the corresponding sequential-generation options when using the API. Prompt guide, API tutorial

A suggested exercise is a sequence showing the same florist preparing flowers, wrapping a bouquet, and placing it on the counter. Describe the persistent subject and setting, then state what differs in each scene.

Evaluate the sequence as a set. Check whether the intended action changes while the subject and environment remain suitable for continuity; do not accept a single strong frame as proof that the whole sequence works.

Keep a compact prompt log with the instruction, references, output, and revision reason. Prefer concrete notes such as missing wrapping paper over vague ratings such as less artistic.

After a few revisions, revisit the brief itself. If you keep changing the desired result, write a new brief rather than comparing it with outputs intended for a different composition.

How do Seedream prompts compare with FLUX.2 prompts?

Seedream 4.0 and 4.5 share an official prompting guide. FLUX.2 also supports generation and editing, with variant-specific controls including adjustable steps and guidance in [flex]. Prompt guide, FLUX documentation

Compare the same creative brief across models while using each service's documented controls. Record model-specific settings separately, so that you can distinguish the instruction from its execution configuration.

The ComfyUI guide provides context for testing prompts inside a larger workflow. Begin with a simple generation stage so later processing does not obscure the effect of a wording change.

What else should you know about Seedream prompts?

How long should a Seedream prompt be?

A Seedream prompt should define the intended scene and the requirements you will inspect. BytePlus recommends concise, precise language rather than repeatedly stacking elaborate vocabulary. Prompt guide

How do I tell Seedream what to preserve?

In a Seedream editing prompt, name the target and describe the elements that should remain unchanged. Use concrete identifiers when several objects or people could match the instruction. Prompt guide

Yes, Seedream 4.0 and 4.5 support image sequences and set-based creation. When using the API, pair that brief with the documented sequential-generation options. Prompt guide, API tutorial

What should I change after an unsuccessful result?

After an unsuccessful Seedream result, choose the most consequential missed requirement and revise that part of the brief. Keep the other instructions stable so you can understand what the revision accomplished.

Sources

Nano Banana Pro: A Practical Guide to ComfyUI API Workflows

Seojun Zhao — Thu, 02 Apr 2026 14:26:25 +0000

To use Nano Banana Pro in ComfyUI, open its official workflow template, sign in to a Comfy account with credits, and supply your prompt and any reference images. The Partner Node sends requests to a hosted service for Google's Gemini 3 Pro Image model. ComfyUI tutorial, Partner Nodes overview, Google announcement

What are the key facts about Nano Banana Pro in ComfyUI?

Fact	Verified detail
Developer	Google DeepMind develops the model; ComfyUI documents its Partner Node integration. Google announcement, ComfyUI tutorial
Released	Model announced November 20, 2025; a separate node release date is not published in the tutorial. Announcement, tutorial
Type	Hosted image generation and editing reached through a ComfyUI Partner Node. Tutorial
Size or parameters	Model parameter count not published; the node is a connector. Model documentation, tutorial
License and access	Closed model with no open weights; Partner Node calls require a Comfy account and credits. Partner Nodes overview
Where it runs	Model inference is remote; the workflow can be opened in local ComfyUI or Comfy Cloud. Tutorial, Partner Nodes overview

Your graph and the model have different execution locations. A local ComfyUI window can send inputs to a remote provider. Account access and the other nodes in your graph therefore deserve separate checks. Partner Nodes overview

What can the Nano Banana Pro ComfyUI workflow do?

ComfyUI's official tutorial provides a ready-made Nano Banana Pro workflow and reference-image inputs.

It documents image blending, text rendering, and subject consistency as model capabilities exposed through the integration. ComfyUI tutorial

For practical work, use that template to keep a reference, instruction, and output in one visible workflow. Start with a simple composition task before adding other stages.

This makes it easier to identify which stage needs attention when the result is wrong.

Google's model documentation also lists thinking and search grounding as supported model capabilities.

Check the node's exposed settings before assuming every Google API capability is available in the same form through ComfyUI. Model documentation

Write a small acceptance brief for the graph. For a product composition, name the product features that must remain recognizable and the intended background.

Use those criteria when comparing returned images, regardless of how many nodes the workflow contains.

What does the Nano Banana Pro Partner Node require?

Partner Nodes require a signed-in Comfy account, credits, and a supported network setup. The official overview explains login under Settings, User, and the credits controls under Settings, Credits. Partner Nodes overview

The provider receives the inputs used by its API node. A workflow containing a hosted node therefore does not keep all generation on your computer, even if you opened ComfyUI locally. Partner Nodes overview

Missing nodes can result from an outdated installation or a failed import. ComfyUI's tutorial warns that templates and node availability can differ between development and stable releases. ComfyUI tutorial

Google's image guide says generated output counts may not exactly follow the request and that generated images contain SynthID.

Inspect actual outputs instead of assuming the request's wording establishes their number or content. Image guide

The model's documentation does not publish a parameter count or local weight package. Its hosted access provides no basis for a fixed local VRAM requirement or a guaranteed generation time for your entire graph. Model documentation

How do you set up Nano Banana Pro in ComfyUI?

Open the official template

Update ComfyUI using its documented installation path. If a required node is missing, check version compatibility and startup import messages. ComfyUI tutorial
Sign in through Settings, User, then inspect the balance under Settings, Credits. Confirm the account and network conditions in the Partner Nodes overview. Overview
Open the Template Library and search for “Nano Banana Pro.” The official tutorial also provides a workflow download and a Comfy Cloud route. Tutorial
Supply your reference images to the template's LoadImage nodes, then enter the requested edit or composition. Inspect the available model settings before running. Tutorial

Keep the template's first run focused. Use a reference of one object and request a different background while preserving the object's shape and colors. This is a suggested diagnostic task, not a claim about measured model performance.

Examine the result before connecting more processing stages. If the request fails, first check whether the problem concerns account access, credits, connectivity, or node availability.

Those are documented prerequisites for the hosted stage. Overview, tutorial

If the request succeeds but the composition is wrong, revise the instruction and references. Treat that as an image-editing problem. Keeping request failures and visual failures separate makes troubleshooting more specific.

Understand the direct API alternative

You can also call the underlying model through Google's Gemini API.

This uses a Google API key and Google's billing context, separately from ComfyUI Partner Node credits. Google announcement, API key guide, Partner Nodes overview

After creating a Gemini API key and setting GEMINI_API_KEY, this minimal request uses the current model identifier and documented API method.

It is a direct Google request, not a command that executes the ComfyUI template. Model documentation, API reference

curl --fail-with-body \
  "https://generativelanguage.googleapis.com/v1beta/models/gemini-3-pro-image:generateContent" \
  -H "x-goog-api-key: $GEMINI_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"contents":[{"parts":[{"text":"Create a studio photograph of a plain red kettle on a cream background."}]}],"generationConfig":{"responseModalities":["TEXT","IMAGE"]}}' \
  -o response.json

The file contains a JSON response envelope. Extract image content from returned parts and handle errors or missing images before passing the result into another stage. API reference

Choose one access route for your initial evaluation and record it with the model name. This keeps account configuration and billing information understandable when someone else revisits the workflow.

For ideas to adapt after the template works, see the sibling Awesome Nano Banana Pro prompt collection guide.

How does Nano Banana Pro compare with Nano Banana 2?

Model	Documented distinction
Nano Banana Pro	Gemini 3 Pro Image focuses on complex visual design, mockups, and text-bearing output. Model documentation
Nano Banana 2	Gemini 3.1 Flash Image is presented as the speed-oriented counterpart for high-volume image work. Nano Banana 2 documentation

Test your own references before choosing a model for a larger graph. Use the ComfyUI complete guide for workflow construction and node connections.

How do you troubleshoot Nano Banana Pro in ComfyUI?

Does the ComfyUI node download Nano Banana Pro?

The Nano Banana Pro ComfyUI integration uses a Partner Node to access a closed hosted model; there are no open Nano Banana Pro weights to load locally. Partner Nodes overview

Do I need a Comfy account for the official Partner Node?

The official Nano Banana Pro Partner Node requires a signed-in Comfy account and available credits. Review the supported network setup as well if login or execution fails. Overview

Can I use my Google API key in the example command?

The Nano Banana Pro curl example calls Google's Gemini API directly with GEMINI_API_KEY. Its billing and authentication are separate from the ComfyUI Partner Node route described above. API key guide, overview

Why is the template missing a node?

Missing Nano Banana Pro workflow nodes can result from ComfyUI version differences or failed startup imports, according to its tutorial. Check the official update guidance and import messages before replacing the node with an unrelated community extension. Tutorial

PromptZone - AI Prompts, Guides and Tools for Builders: Seojun Zhao

Is 30% of arXiv Now AI-Written?

Scale of the Shift

Detection Method

Key Numbers Reported

Detection Tools Compared

Who Should Track This

Practical Next Steps

GPT-5.6 Edges Grok 4.5 in App Build-Off

Build-Off Task Breakdown

Output Metrics from the Thread

How to Replicate the Test

Observed Tradeoffs

Who Should Run Similar Tests

Model Selection Verdict

Pulpie Delivers Pareto-Optimal Web Cleaning Models

What It Is and How It Works

Benchmarks and Key Numbers

How to Try It

Pros and Cons

Alternatives and Comparisons

Who Should Use This

HiDream-I1 License Guide to Open Weights and Commercial Use

What are the key facts about HiDream-I1 licensing?

What does the HiDream-I1 release let you use?

Which HiDream-I1 components have separate license terms?

How do you evaluate HiDream-I1 for a commercial project?

How do HiDream-I1 and FLUX.1-schnell licenses compare?

What should you know before adopting HiDream-I1?

What license covers HiDream-I1 code and weights?

Can I use HiDream-I1 images commercially?

Do Full, Dev, and Fast use different licenses?

Does the 17B parameter count include all deployment components?

Sources

Related guides on PromptZone

Seedream Prompts: A Practical Guide to Scenes and Image Edits

What are the key facts about Seedream 4.0 and 4.5 prompts?

What should you include in a Seedream prompt?

What can a Seedream prompt control?

How do you write and test Seedream prompts?

How do Seedream prompts compare with FLUX.2 prompts?

What else should you know about Seedream prompts?

How long should a Seedream prompt be?

How do I tell Seedream what to preserve?

Can I request a related set of images?

What should I change after an unsuccessful result?

Sources

Related guides on PromptZone

Nano Banana Pro: A Practical Guide to ComfyUI API Workflows

What are the key facts about Nano Banana Pro in ComfyUI?

What can the Nano Banana Pro ComfyUI workflow do?

What does the Nano Banana Pro Partner Node require?

How do you set up Nano Banana Pro in ComfyUI?

Open the official template

Understand the direct API alternative

How does Nano Banana Pro compare with Nano Banana 2?

How do you troubleshoot Nano Banana Pro in ComfyUI?

Does the ComfyUI node download Nano Banana Pro?

Do I need a Comfy account for the official Partner Node?

Can I use my Google API key in the example command?

Why is the template missing a node?

Sources

Related guides on PromptZone