AI Guides › Step-by-step guides
By Nigel Guy · 8 min read
The usual attempt: upload one good selfie, ask for "me in a linen suit on a rooftop", and the first image is close enough to be exciting. By the sixth the jaw has moved, the eyes have changed shape and the set no longer looks like one person, let alone you. More prompting doesn't fix that. Training the face once, then shooting every image from that trained identity, does.
The rule: train the face on the Higgsfield website, direct the shoot from Claude, and generate every image from the trained character by name. Never use your last output as the reference for your next one.
Image models treat each request on its own. Higgsfield's help centre makes the distinction plainly: a reference image anchors a single generation but drifts across many, while Soul ID trains a reusable identity once and applies it to later generations. Higgsfield also sets the bar honestly: expect "clearly the same person", not a pixel-identical face, and extreme styles or unusual angles can still introduce small drift.
Claude connects to Higgsfield's remote MCP server, which Claude's menus call a custom connector. Higgsfield's connector page lists Soul characters, reference Elements, image and video generation, upscaling and credit checks as available through it. Training a new Soul ID through the connector is not listed, so this guide does the training on higgsfield.ai and uses the connector only for the shoot.
| You need | Cost at time of writing | Notes |
|---|---|---|
| A Claude account (claude.ai or Claude Desktop) | Free plan works; Pro is listed at about £18 a month in the UK | Free accounts are limited to one custom connector |
| A paid Higgsfield subscription | Billed in US dollars; check the pricing page before buying | Higgsfield says an active paid subscription is required for the connector |
| 20 to 80 recent photos of yourself | — | Different angles and expressions, at least one full-length shot |
| A list of 6 to 12 shots you want | — | One look or theme per set |
Two cost traps. First, Higgsfield's help centre says every generation made through a connected agent deducts credits at standard rates, and "unlimited" access and free generations apply only on higgsfield.ai itself. Second, there is no built-in spending cap over the connector; any limit you set is an instruction in the chat, not a hard stop. Higgsfield's own blog lists Soul ID image generations at 25 credits each at time of writing. I could not confirm whether training itself costs credits, so check your balance before and after.
Pick 20 or more photos taken in the last four or five months. Higgsfield's guidance: varied angles and expressions, clear even light, one person per photo, no sunglasses, hats or scarves over the face, no cropped faces, at least one full-height photo. Leave out filtered or beauty-smoothed images. Attach a batch to a Claude chat (Claude can see images you attach, even though Higgsfield can't read them from there) and use this prompt. Fill in the bracketed parts.
You are a portrait photographer reviewing a training set for an AI identity model. The goal is a set that teaches the model my face under varied, honest conditions.
Context: I am training a Higgsfield Soul ID. It needs at least 20 photos of one person, up to 80. My current appearance: [ANY_RECENT_CHANGES, e.g. shorter hair since March, now wear glasses].
For each attached photo, in order:
1. Give it a number and a one-word verdict: KEEP, MAYBE or DROP.
2. If MAYBE or DROP, give the reason in under ten words (shadow, sunglasses, cropped face, filter, group shot, outdated look, near-duplicate pose).
Then give me:
- A count of KEEPs.
- What the set is missing (angles, expressions, a full-length shot, lighting variety).
- Whether the KEEP set is ready to train, yes or no.
Do not guess at photos you cannot see clearly; say so. If I have attached fewer than 20 photos, tell me how many more I need before judging readiness.
Before you answer, check: did you flag every near-duplicate and every filtered image?
https://mcp.higgsfield.ai/mcp.You are the art director on a personal photoshoot. The subject is my trained Higgsfield Soul character named [CHARACTER_NAME]. The aim is a set of [NUMBER_OF_IMAGES] photorealistic images that read as one shoot of one person.
Brief:
- Use: [USE, e.g. LinkedIn header, podcast artwork, website about page]
- Look and wardrobe: [WARDROBE_AND_STYLE]
- Locations or backdrops: [LOCATIONS]
- Aspect ratios needed: [RATIOS, e.g. 4:5 and 16:9]
- Credit budget for the whole set: [MAX_CREDITS]
Steps:
1. Write a numbered shot list: framing, pose, expression, setting, lighting. Keep wardrobe and colour palette constant unless I say otherwise.
2. Avoid shots that the identity model handles poorly: extreme profiles, faces hidden by hands or props, heavy stylisation. Flag any shot in my brief that falls into those.
3. Estimate the credit cost per image if the Higgsfield tools report it, and the total.
4. Stop and wait for my approval. Do not generate anything in this step.
If any brief item is missing or vague, ask me rather than filling it in.
Before replying, check: does every shot name [CHARACTER_NAME] as the subject, and have you stopped before spending credits?
Once the shot list is approved, send this:
Generate shots [SHOT_NUMBERS] from the approved list with Higgsfield, using my Soul character [CHARACTER_NAME] as the identity in every image.
Rules:
- Tell me the credit cost and wait for my go-ahead before each batch.
- Do not use any previously generated image as a face reference. The trained character is the only identity source.
- Do not exceed [MAX_CREDITS] credits in total; stop and tell me if the next batch would.
- After each batch, list the images by shot number so I can approve or reject them.
Work in batches of two or three.
Shot [SHOT_NUMBER] has drifted: [WHAT_CHANGED, e.g. nose wider, hair darker, looks older].
Discard it. Regenerate the same shot from my Soul character [CHARACTER_NAME] only, with the same scene description and a simpler pose or a more front-facing angle. Give me the cost first.
If one shot drifts twice, the shot is the problem; change the angle or drop it. If most shots drift, the training set is the problem; go back to Step 1 and retrain on a stronger set.