Methodology
How this wiki was made, what its labels mean, and where it can be wrong.
Sources
- Videos: the uploads of Higgsfield AI's YouTube channel (long-form, Shorts, livestreams) and the lessons of Higgsfield Academy: 533 recordings recordings (41 hours), 2025–2026.
- Texts: 666 texts Higgsfield blog articles, changelog entries, help-centre articles and tool, model, MCP and app pages. Tool and app pages describe the product on the day they were saved ("as of"), not a release date.
- Comments: the channel's own comments, and viewer questions the channel answered (the question is shown as the context of the answer; viewers' words are never recorded as the channel's).
- Transcripts: Whisper large-v3; the channel uploads no captions. Model and tool names are normalised to their official spelling; misheard names may remain in quotes.
How each source was read
- An analyst model reads the transcript or text and records statements (how-to steps, recommendations, capabilities, limitations, prompt patterns, settings, prices, workflow stages, comparisons, studio practice, releases), the models, tools and techniques named, and dated answers to practice questions.
- A second model audits that record against the source and fixes invented or distorted content, wrong attribution, marketing read as fact, and missed answers.
- Code checks that every quote and evidence span is copied from the source and anchors it to its time.
- A registrar merges new models, tools, techniques and threads into one registry.
Nothing is added from outside knowledge.
Labels
- Vendor claim: a marketing statement by Higgsfield or a model maker ("no identity drift", "15× cheaper"). Kept, never used as an answer on its own.
- One creator's experience: what worked for one creator on one project with one model version.
- Time-bound: an offer or deadline, shown with its end date and left out of evergreen answers.
Practice questions
Each answer is dated and attributed: first on record, restate, refine, shift, revise, tension
Threads and their statuses
A thread follows one thing across sources. Its status is a state, not the last thing that happened:
| kind | status |
|---|---|
| release history | proposed, running, completed, lapsed; completed, not continued; last documented in YYYY when nothing new is recorded for over two years; known from looking back |
| promise | kept, kept late, dropped, changed; the project it belongs to reported its result; standing commitment; open (no report in 2025–2026); open (updates through YYYY, not reported done); postponed (YYYY); not reported kept; said in jest (not tracked) |
| ask | reported back; the project it belongs to reported its result; asked again year after year; asked again in N years (YYYY–YYYY); asked in YYYY; no response reported; reminder only; promotion, announcement, invitation, joke or advice (not a request) |
| prediction | came true, partly came true, proved wrong, or unresolved (no horizon named); due YYYY; outcome not yet known; its date passed (YYYY); no outcome reported; said in jest (not tracked) |
| story | told N×; followed over N videos; updates only |
The events on a thread are named:
| kind | events |
|---|---|
| ask | asked, reminder, result reported, thanks, closed |
| prediction | predicted, repeated, recalled, came true, partly came true, proved wrong |
| promise | promised, progress update, kept, dropped, changed, kept late, postponed, recalled |
| project history | proposed, announced, launched, milestone, setback, rule change, paused, restarted, outcome, looked back |
| story | told, retold, update, mentioned, stopped |
Dates
Dates are upload dates, which set the order. Where the speaker gives a different recording date in the video, it is added as "said on YYYY-MM-DD, implied" (derived from the date said in the video).
Limits
- Machine reading makes mistakes; the audit catches many, not all.
- Coverage ends October 24, 2026: later releases and changes in advice are not here.
- Prices, limits and plans change often; every figure carries its date.